A renewable energy power generation power prediction and power dispatch method and system

Abstract

The application discloses a renewable energy power generation power prediction and power dispatching method and system, comprising: collecting historical power generation data and real-time meteorological data of renewable energy power generation, performing linear interpolation on the historical power generation data and real-time meteorological data to fill in missing values and box plot anomaly detection, and obtaining a normalized training data set; using the normalized training data set, constructing a hybrid prediction model by using a neural symbol acceleration technology with time logic constraints, extracting medium and long term spatial and temporal features, and generating a renewable energy power generation power prediction result; according to the renewable energy power generation power prediction result and system constraint conditions, adopting a linear one-dimensional projection constrained distribution robust control method to formulate a dispatching strategy, and using the dispatching strategy to solve an optimal dispatching scheme by a mixed integer linear programming. The application significantly improves the renewable energy power generation power prediction accuracy and the robustness of power dispatching.

Description

Technical Field

[0001] This application relates to the field of power systems, and in particular to a method and system for predicting renewable energy power generation and dispatching power, which is used to achieve high-precision prediction of renewable energy power generation and optimized dispatching of power systems based on the prediction results. Background Technology

[0002] Renewable energy power generation forecasting and power dispatching technologies are key technologies for achieving a clean energy transition and ensuring the safe and stable operation of power systems. With the increasing penetration of renewable energy sources such as wind and solar power in power systems, accurately forecasting their power generation and formulating reasonable dispatching strategies has become a major challenge for power system operation.

[0003] Existing renewable energy power prediction technologies mainly include physical modeling methods and statistical learning methods. Physical modeling methods predict power by establishing mathematical relationships between wind speed and power or solar radiation and power, but they struggle to handle complex nonlinear relationships. Statistical learning methods, such as support vector machines and random forests, can capture nonlinear features in the data, but they have limitations in handling long-term time-series dependencies and multivariate coupling. In recent years, deep learning methods, such as LSTM and CNN, have been widely used in power prediction, effectively extracting time-series features.

[0004] In power dispatching, traditional deterministic optimization methods typically formulate dispatching strategies based on point forecasts and employ linear programming or quadratic programming to solve for the optimal dispatching scheme. These methods assume that the forecasts are completely accurate and ignore the randomness and uncertainty of renewable energy output. Although stochastic optimization and robust optimization methods can handle uncertainty to some extent, they still have shortcomings in terms of computational complexity and solution efficiency.

[0005] However, existing technologies have two key problems: first, the prediction models lack effective handling of time logic constraints, resulting in defects in the physical feasibility of the prediction results; second, the scheduling strategies are not robust enough to prediction uncertainties, making it difficult to guarantee the safe operation of the system in extreme scenarios. Summary of the Invention

[0006] Therefore, the purpose of this invention is to provide a method and system for predicting renewable energy power generation and dispatching power, which aims to solve the problems of insufficient time logic constraint processing in the prediction model and insufficient robustness of the dispatching strategy in the prior art, and to achieve high-precision prediction of renewable energy power generation and optimized dispatching of the power system based on the prediction results.

[0007] To achieve the above objectives, the present invention provides a method for predicting renewable energy power generation and dispatching electricity, comprising the following steps:

[0008] Historical power generation data and real-time meteorological data of renewable energy generation are collected. Linear interpolation is used to fill missing values ​​and box plot anomalies are detected on the historical power generation data and real-time meteorological data to obtain a normalized training dataset.

[0009] Using the normalized training dataset, a hybrid prediction model is constructed using a neural symbolic acceleration technique with temporal logic constraints. The hybrid prediction model includes an LSTM network and a CNN-Transformer network. The LSTM network captures short-term temporal dependencies, and the CNN-Transformer network extracts medium- and long-term spatial-temporal features to generate renewable energy power generation prediction results.

[0010] Based on the renewable energy power generation prediction results and system constraints, a scheduling strategy is formulated using a linear one-dimensional projection constraint sub-Blu-rod control method, and the optimal scheduling scheme is solved by mixed integer linear programming using the scheduling strategy.

[0011] Based on the optimal scheduling scheme, a push-pull average consensus algorithm with dynamic compression is used to achieve multi-device coordinated control. An adaptive quantization scheme is used to achieve scheduling consensus and execute real-time power adjustment operations.

[0012] Based on the actual execution status of the equipment and the operating status of the system, the hybrid prediction model and scheduling strategy are continuously optimized through reinforcement learning algorithms to form an adaptive closed-loop control system.

[0013] Preferably, historical power generation data and real-time meteorological data of renewable energy generation are collected, and linear interpolation is performed on the historical power generation data and real-time meteorological data to fill missing values ​​and box plot anomaly detection is performed to obtain a normalized training dataset, including:

[0014] The historical power generation data of wind power and photovoltaic power stations and real-time meteorological data of meteorological observation stations are collected by the SCADA system. The consistency of the timestamps of the historical power generation data and the real-time meteorological data is verified to obtain synchronized raw data.

[0015] The original synchronized data is subjected to quality checks to identify missing data, outliers, and duplicate records. Missing values ​​are filled using linear interpolation, and outliers are identified using box plots to generate cleaned data.

[0016] The cleaned data is subjected to Min-Max normalization to map the data to the interval between 0 and 1. Periodic features of hours, days of the week, and months are extracted and lagged features are generated to obtain the normalized training dataset.

[0017] Preferably, missing values ​​are filled using linear interpolation and outliers are identified using box plots to generate cleaned data, including:

[0018] Missing value detection is performed on the synchronized raw data. When the continuous missing time is less than 2 hours, the missing value is calculated based on the data before and after the time. When the continuous missing time is more than 2 hours, the missing value is replaced by historical data from the same period to generate the completed data.

[0019] Based on the completed data, the first quartile, third quartile, and interquartile range of each parameter are calculated. Outlier data points exceeding 1.5 times the interquartile range are identified using the box plot method, and outlier labeling results are generated.

[0020] Abnormal data points are removed based on the outlier marking results. The integrity of the removed data is then verified to ensure data continuity and consistency, resulting in the cleaned data.

[0021] Preferably, capturing short-term temporal dependencies using an LSTM network includes:

[0022] Based on the normalized training dataset, a multi-layer LSTM network is constructed with a hidden layer dimension of 128, and information from key time steps is adaptively selected through an attention mechanism.

[0023] The power generation, meteorological data and equipment status information of the past 24 hours are input into the LSTM network for processing, short-term time-series dependency features are extracted, and hourly power prediction values ​​for the next six hours are generated.

[0024] Introducing time logic constraints ensures that the predicted power change rate does not exceed the equipment ramp-up capability, thus obtaining short-term power prediction results that satisfy physical constraints.

[0025] Preferably, the CNN-Transformer network extracts medium- to long-term spatial-temporal features, including:

[0026] Based on the normalized training dataset, a CNN-Transformer hybrid architecture is designed. The CNN module uses one-dimensional convolution to extract local temporal patterns, and the convolution kernel sizes are 3, 5, and 7.

[0027] By capturing the changing patterns of different cycles through multi-scale feature fusion, the Transformer module uses a multi-head self-attention mechanism to handle long-term dependencies, with the number of heads set to 8 and the feedforward network dimension to 512.

[0028] By using neural symbolic acceleration technology, the prediction problem is modeled as a constrained optimization problem. The approximate optimal solution is quickly solved by neural network, and the results are verified and corrected by symbolic reasoning to obtain medium- and long-term power prediction results that meet the physical feasibility constraints.

[0029] Preferably, based on the renewable energy power generation forecast results and system constraints, a scheduling strategy is formulated using a linear one-dimensional projective constraint sub-Bruker control method, and the optimal scheduling scheme is solved using mixed-integer linear programming with the scheduling strategy, including:

[0030] Based on the renewable energy power generation prediction results, a multi-objective scheduling model that comprehensively considers economic efficiency and security is established. The weighted sum method is used to unify minimizing power purchase cost, maximizing renewable energy absorption rate, and minimizing system frequency deviation into a single optimization objective.

[0031] Construct a complete system of constraints, including power balance constraints, equipment operation constraints, network security constraints, and unit ramp-up constraints, to ensure the feasibility and safety of the dispatching scheme;

[0032] The uncertainty of renewable energy output is handled by a linear one-dimensional projection constraint sub-Bruker control method. The uncertainty set is constructed by using the difference between the actual distribution and the empirical distribution through the Wasserstein distance metric, and the optimal scheduling scheme is obtained.

[0033] Preferably, a linear one-dimensional projective constraint-based blobs control method is used to handle the uncertainty of renewable energy output. An uncertainty set is constructed by using the Wasserstein distance metric to measure the difference between the true and empirical distributions, including:

[0034] Based on the renewable energy power generation prediction results, the uncertainty set is defined as a confidence interval centered on the predicted value. The uncertainty interval for wind power is set to ±25% of the predicted value, and the uncertainty interval for photovoltaic power is set to ±20% of the predicted value, thus generating an initial uncertainty set.

[0035] The difference between the true distribution and the empirical distribution of the initial uncertainty set is measured by the Wasserstein distance. The Wasserstein radius is set to 0.1. Based on the statistical characteristics of historical prediction errors, the confidence level is determined to be 95%, and a split-bar uncertainty set is generated.

[0036] In the set of uncertainties in the sub-Brussels bar, the sub-Brussels bar optimization problem is transformed into a deterministic mixed-integer linear programming problem using duality theory. This eliminates the complexity of probabilistic constraints and ensures that the system's safe operation requirements are still met even in the worst case, thus obtaining the optimal scheduling scheme.

[0037] Preferably, based on the optimal scheduling scheme, a push-pull average consensus algorithm with dynamic compression is used to achieve multi-device coordinated control, and a scheduling consensus is achieved through an adaptive quantization scheme to perform real-time power adjustment operations, including:

[0038] Based on the optimal scheduling scheme, the energy storage system, controllable load, and distributed power source in the power system are modeled as intelligent agents with autonomous decision-making capabilities, and each agent is equipped with an independent computing unit and communication module.

[0039] Design a distributed protocol based on the push-pull average consensus algorithm, in which each agent periodically exchanges information with a predefined set of neighbor agents, wherein the neighbor relationship is determined based on the physical connection topology and communication network structure;

[0040] During the push operation phase, each agent sends its current state vector and gradient information to all neighboring agents, and the sending weights are dynamically adjusted according to the degree centrality of the network topology.

[0041] During the pull operation phase, each agent receives state information from its neighbors and updates its local state by merging the neighbor information using a weighted average method.

[0042] A multi-level convergence criterion is established, and scheduling consensus is reached through state convergence, objective function convergence, and consistency error criterion to execute the real-time power adjustment operation.

[0043] Preferably, the adaptive quantization scheme includes:

[0044] The information-theoretic-based dynamic compression mechanism assesses the importance of information based on Shannon entropy and adaptively adjusts the quantization precision to reduce the communication burden.

[0045] A three-level quantization system is established, with key control signals using 16-bit high-precision quantization, status monitoring information using 8-bit medium-precision quantization, and auxiliary information using 4-bit low-precision quantization.

[0046] The system monitors the variance and rate of change of the system's operating state in real time. When the system is in a steady state, a coarse quantization strategy is adopted to reduce communication overhead. When a disturbance is detected in the system's operating state, the system automatically switches to fine quantization mode to ensure control accuracy.

[0047] Preferably, based on the actual execution state of the equipment and the operating state of the system, the hybrid prediction model and scheduling strategy are continuously optimized through reinforcement learning algorithms to form an adaptive closed-loop control system, including:

[0048] Establish a real-time data feedback channel to collect the actual operating status of the equipment through the SCADA system and smart meters, including the actual charging and discharging power of the energy storage system, the actual adjustment of the controllable load, and the actual output of renewable energy.

[0049] Calculate prediction error metrics, which include mean absolute error, root mean square error and mean absolute percentage error. When any metric exceeds a preset threshold, a model retraining mechanism is triggered.

[0050] The online learning algorithm is designed to update the prediction model parameters using an incremental learning approach. The weight of the updated sample is set to 0.1, and the weight of the historical sample is adjusted according to an exponential decay law.

[0051] Construct a scheduling performance evaluation system, including economic indicators, safety indicators, and environmental indicators. When the performance indicators deviate from the target value by more than 5%, trigger an adaptive adjustment mechanism for the scheduling strategy.

[0052] Based on the aforementioned scheduling performance evaluation system, the system operating state is continuously optimized using a reinforcement learning algorithm, with the scheduling action as the output and a multi-objective weighted sum as the reward function, thus forming the adaptive closed-loop control system.

[0053] Preferably, the reinforcement learning algorithm includes:

[0054] Based on the aforementioned scheduling performance evaluation system, the scheduling strategy is continuously optimized using a deep deterministic policy gradient algorithm, taking the system running state as input, the scheduling action as output, and a multi-objective weighted sum as the reward function.

[0055] Construct a state space, including renewable energy forecasting bias, energy storage state of charge, load change rate, and electricity price information; and an action space, including energy storage charging and discharging commands, load adjustment commands, and grid interaction power.

[0056] Set up an experience replay buffer to store historical state-action-reward sequences. The buffer size is 10,000. Perform batch learning every 100 steps with a batch size of 32.

[0057] An actor network and a critic network are established, wherein the actor network outputs continuous action values, and the critic network evaluates the value function of state-action pairs;

[0058] The actor network parameters are updated using the gradient ascent method to maximize the cumulative reward, and the critic network parameters are updated using the temporal difference error method to form a continuously optimized reinforcement learning strategy.

[0059] Preferably, a hybrid prediction model is constructed using a neural symbolic acceleration technique with time logic constraints, including:

[0060] Based on the normalized training dataset, the power change rate constraint and equipment capacity constraint are encoded as symbolic rules, and the constraint rules are embedded into the deep learning model through the neural symbolic interface layer.

[0061] A time logic constraint processing module is constructed to encode the constraint that the wind power change rate does not exceed 10% of the rated power per minute and the photovoltaic power change rate does not exceed 20% of the rated power per minute, and generate constraint embedding vectors.

[0062] The prediction problem is modeled as a constrained optimization problem using the neural symbol acceleration technique of MILP motion planning. The approximate optimal solution is quickly solved by a neural network, and the prediction results are verified and corrected by symbolic reasoning to obtain the hybrid prediction model that satisfies the time logic constraints.

[0063] This invention also provides a renewable energy power generation prediction and power dispatch system, comprising:

[0064] The data acquisition and preprocessing module is used to acquire historical power generation data and real-time meteorological data of renewable energy generation based on a standardized data interface, and to perform linear interpolation to fill missing values ​​and box plot anomaly detection on the historical power generation data and real-time meteorological data to obtain a normalized training dataset.

[0065] The neural symbol fusion prediction module is used to construct a hybrid prediction model based on the normalized training dataset using neural symbol acceleration technology with temporal logic constraints. It captures short-term temporal dependencies through an LSTM network and extracts medium- and long-term spatial-temporal features through a CNN-Transformer network to generate renewable energy power generation prediction results.

[0066] The sub-Blu-bar scheduling optimization module is used to formulate a scheduling strategy based on the renewable energy power generation prediction results and system constraints, using a linear one-dimensional projection constraint sub-Blu-bar control method, and to solve the optimal scheduling scheme through mixed integer linear programming.

[0067] The distributed coordination and control module is used to achieve multi-device coordination and control based on the optimal scheduling scheme, using a push-pull average consensus algorithm with dynamic compression, achieving scheduling consensus through an adaptive quantization scheme, and performing real-time power adjustment operations.

[0068] The closed-loop feedback optimization module is used to continuously optimize the hybrid prediction model and scheduling strategy based on the actual execution status of the equipment and the operating status of the system, thereby forming an adaptive closed-loop control system.

[0069] Compared with the prior art, the present invention has the following beneficial effects:

[0070] This invention is the first to apply MILP motion planning neural symbol acceleration technology with time logic and chance constraints to renewable energy power prediction. By encoding physical constraints into symbolic rules and embedding them into neural networks, the prediction accuracy and computational efficiency are significantly improved, and the prediction error is reduced by more than 15%.

[0071] This invention innovatively employs a linear one-dimensional projection constraint-based bibliometric control method to handle the uncertainty of renewable energy output. By constructing an uncertainty set through Wasserstein distance, the robustness of the scheduling strategy under various scenarios is guaranteed, and the system security is improved by 20%.

[0072] This invention introduces a push-pull average consensus algorithm with dynamic compression (PP-ACDC) to achieve efficient coordination among distributed devices. It achieves optimal scheduling consensus without the need for global information through an adaptive quantization scheme, reducing communication overhead by 30%.

[0073] This invention constructs a unified optimization framework with multiple time scales, from minute-level prediction to daily-level scheduling, to achieve an organic combination of accurate short-term prediction and robust long-term scheduling, thereby improving the overall system operating efficiency by 25%. Attached Figure Description

[0074] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to this disclosure and, together with the specification, serve to explain the technical solutions of this disclosure. It should be understood that the following drawings only show some embodiments of this disclosure and should not be considered as limiting the scope. Those skilled in the art can obtain other related drawings based on these drawings without creative effort.

[0075] Figure 1 This is a flowchart of the renewable energy power generation prediction and power dispatch method of the present invention;

[0076] Figure 2 This is a block diagram of the renewable energy power generation prediction and power dispatch system of the present invention. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. The components of the embodiments of this disclosure described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed disclosure, but merely represents selected embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.

[0078] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0079] In this document, the term "and / or" merely describes a relationship, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0080] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0081] like Figure 1 As shown, the present invention provides a method for predicting renewable energy power generation and dispatching electricity, comprising the following steps:

[0082] Step 1: Multi-source data acquisition and intelligent preprocessing;

[0083] Historical power generation data and real-time meteorological data of renewable energy generation are collected. Linear interpolation is used to fill missing values ​​and box plot anomalies are detected on the historical power generation data and real-time meteorological data to obtain a normalized training dataset.

[0084] In step one, multi-source data acquisition and intelligent preprocessing are the fundamental steps of this invention. It collects historical power generation data from wind and photovoltaic power plants through a SCADA system, including operating parameters such as active power, reactive power, voltage, and current, with a sampling frequency of 15 minutes. Simultaneously, it collects real-time meteorological data from meteorological observation stations, including wind speed, wind direction, temperature, humidity, air pressure, and light intensity, ensuring the consistency of data timestamps. The collected raw data undergoes quality checks to identify missing data, outliers, and duplicate records. When the continuous missing time is less than 2 hours, linear interpolation is used to fill in the missing values; when the continuous missing time exceeds 2 hours, historical data from the same period is used as replacements. Outliers are identified using box plots, and data points exceeding 1.5 times the interquartile range are marked as outliers and removed. Finally, all numerical data undergoes Min-Max normalization, mapping the data to the [0,1] interval to eliminate the influence of different units on model training, and periodic features such as hours, days of the week, and months are extracted to generate lag features to enhance time-series modeling capabilities.

[0085] Step 2: Construction of a neural symbol fusion-based power generation prediction model;

[0086] Using the normalized training dataset, a hybrid prediction model is constructed using a neural symbolic acceleration technique with temporal logic constraints. The hybrid prediction model includes an LSTM network and a CNN-Transformer network. The LSTM network captures short-term temporal dependencies, and the CNN-Transformer network extracts medium- and long-term spatial-temporal features to generate renewable energy power generation prediction results.

[0087] Step two, the construction of the neural symbol fusion power generation prediction model, is the core technical step of this invention. This step designs a neural symbol architecture with temporal logical constraints, encoding physical constraints such as power change rate limits and equipment capacity constraints into symbolic rules. These constraints are then embedded into the deep learning model through a neural symbol interface layer. For short-term predictions (0-6 hours), a multi-layer LSTM network is constructed with a hidden layer dimension of 128. An attention mechanism adaptively selects information from key time steps. Input features include power generation, meteorological data, and equipment status information from the past 24 hours, ensuring that the predicted power change rate does not exceed the equipment's ramp-up capability (10% / minute of rated power for wind power, and 20% / minute for photovoltaic power). For medium- to long-term predictions (1-7 days), a CNN-Transformer hybrid architecture is designed. The CNN module uses 1D convolutions to extract local temporal patterns, with kernel sizes of 3, 5, and 7. Multi-scale feature fusion captures the changing patterns of different cycles. The Transformer module uses a multi-head self-attention mechanism to handle long-term dependencies, with 8 heads and a feedforward network dimension of 512. Finally, the neural symbolic acceleration technology of MILP motion planning is used to model the prediction problem as a constrained optimization problem. The approximate optimal solution is quickly solved by neural network, and the results are verified and corrected by symbolic reasoning to ensure that the predicted value meets the physical feasibility constraints.

[0088] Step 3: Optimize the scheduling strategy of the distributed bar;

[0089] Based on the renewable energy power generation prediction results and system constraints, a scheduling strategy is formulated using a linear one-dimensional projection constraint sub-Blu-rod control method, and the optimal scheduling scheme is solved by mixed integer linear programming using the scheduling strategy.

[0090] The linear one-dimensional projective constraint-based bibloc control method is an advanced optimization technique for handling uncertainties in renewable energy output. This method constructs an uncertainty set to characterize the difference between the actual and empirical distributions, ensuring system safety even in the worst-case scenario. The core idea is to project complex distribution information into a one-dimensional space for processing, significantly reducing computational complexity while maintaining the robustness of the optimization results.

[0091] This method first constructs an empirical distribution of renewable energy output based on historical data, and then uses Wasserstein distance to measure the deviation between the actual distribution and the empirical distribution. Wasserstein distance effectively captures structural differences between distributions and offers better geometric intuition and computational stability compared to traditional probabilistic distance metrics. By setting an appropriate Wasserstein radius, the system can construct an uncertainty set encompassing all possible actual distributions, ensuring the feasibility of optimization results under various distribution conditions. Linear one-dimensional projection technology maps high-dimensional uncertainty information to one-dimensional space, retaining the most critical uncertainty features by selecting an appropriate projection direction, thus reducing computational burden and avoiding information loss.

[0092] Step three includes:

[0093] In the data preprocessing stage, the system collects historical renewable energy output data, cleans and standardizes the data, identifies and removes outliers, and ensures the representativeness and accuracy of the empirical distribution. In the uncertainty modeling stage, the system constructs an empirical distribution based on the cleaned historical data and calculates the distribution's statistical characteristics, including key indicators such as mean, variance, skewness, and kurtosis. The system determines the Wasserstein radius based on the historical statistical characteristics of the prediction error, typically set to cover 95% of historical prediction error scenarios. In the projection direction selection stage, the system uses principal component analysis to identify the main directions of uncertainty change and selects the principal components with the largest variance contribution as projection directions, ensuring that the projected one-dimensional information fully represents the original high-dimensional uncertainty.

[0094] In the construction phase of the subbulbar optimization model, the system transforms the original stochastic optimization problem into a subbulbar optimization problem. The objective function is designed to minimize the expected cost under the worst-case distribution, and the constraints include safe operation constraints that can be satisfied under all possible distributions. The system uses duality theory to transform the infinite-dimensional subbulbar optimization problem into a finite-dimensional deterministic optimization problem. By introducing dual variables, the inner-layer maximization problem and the outer-layer minimization problem of the original problem are exchanged, eliminating the complexity of probabilistic constraints.

[0095] In the mixed-integer linear programming solution stage, the system further organizes the transformed deterministic optimization problem into a standard mixed-integer linear programming form. The system performs piecewise linearization on the nonlinear constraints, introducing binary variables to represent the start / stop status and operating mode selection of the equipment, and continuous variables to represent the output level and adjustment amount of the equipment. The system uses a professional commercial solver for solving the problem, setting appropriate solution accuracy and time limits to ensure that high-quality solutions are obtained within the time window required by real-time scheduling. To improve solution efficiency, the system also implements a hot-start mechanism, using the optimization results of the previous time period as the initial solution for the current time period to accelerate the solution process.

[0096] During the results verification and adjustment phase, the system comprehensively verifies the obtained scheduling schemes, checking whether they meet all physical constraints and safety requirements, and evaluating the economic and environmental indicators of the schemes. When a scheme fails to meet certain requirements, the system automatically adjusts the constraints or optimization parameters and re-solves until satisfactory results are obtained. The system also establishes a sensitivity analysis mechanism for scheduling schemes to assess the impact of changes in key parameters on the optimization results, providing decision support information for operators.

[0097] For example, a multi-objective scheduling model that comprehensively considers economics and security is first established. A weighted sum method is used to unify the three core indicators—minimizing electricity purchase cost, maximizing renewable energy absorption rate, and minimizing system frequency deviation—into a single optimization objective, with weighting coefficients set to 0.5, 0.3, and 0.2, respectively. Electricity purchase cost is calculated using time-of-use pricing, divided into peak, normal, and valley periods based on grid load characteristics. A complete system constraint framework is constructed, including power balance constraints, equipment operation constraints, network security constraints, and unit ramp-up constraints, ensuring the feasibility and security of the scheduling scheme. A sub-Brook bar control method with linear one-dimensional projection constraints is used to handle the uncertainty of renewable energy output. The uncertainty set is defined as a confidence interval centered on the predicted value; the uncertainty interval for wind power is ±25% of the predicted value, and for photovoltaics, it is ±20%. The difference between the true and empirical distributions is measured using Wasserstein distance, and duality theory is used to transform the sub-Brook bar optimization problem into a deterministic mixed-integer linear programming problem, ensuring that the system's safe operation requirements are met even in the worst-case scenario.

[0098] Step 4: Distributed real-time scheduling based on consensus algorithms;

[0099] Based on the optimal scheduling scheme, a push-pull average consensus algorithm with dynamic compression is used to achieve multi-device coordinated control. An adaptive quantization scheme is used to achieve scheduling consensus and execute real-time power adjustment operations.

[0100] Step four introduces a distributed real-time scheduling system based on a consensus algorithm to achieve coordinated control among multiple distributed devices. Energy storage systems, controllable loads, and distributed power sources in the power system are modeled as intelligent agents with autonomous decision-making capabilities. Each agent is equipped with an independent computing unit and communication module, maintaining local state information including operating parameters such as the energy storage system's SOC, charge / discharge efficiency, cycle count, and temperature. A distributed protocol based on the push-pull average consensus algorithm (PP-ACDC) is designed, with each agent periodically exchanging information with a predefined set of neighboring agents. During the push operation phase, the agent sends its current state vector and gradient information to all neighboring agents; during the pull operation phase, the agent receives state information from its neighbors and updates its local state by merging neighbor information through a weighted average. A dynamic compression mechanism based on information theory is introduced, establishing a three-level quantization system: key control signals use 16-bit high-precision quantization, state monitoring information uses 8-bit medium-precision quantization, and auxiliary information uses 4-bit low-precision quantization. The variance and rate of change of the system's operating state are monitored in real time. When the system is in a steady state, a coarse quantization strategy is used to reduce communication overhead; when a system disturbance is detected, it automatically switches to fine quantization mode to ensure control accuracy.

[0101] The push-pull average consensus algorithm with dynamic compression is an advanced distributed coordination and control technology. This algorithm achieves globally optimal control through local information exchange between intelligent agents, without the need for a central controller. Its core advantage lies in combining push and pull information propagation mechanisms, which accelerates the convergence process and reduces communication complexity, making it particularly suitable for applications like power systems with extremely high real-time requirements.

[0102] Furthermore, the push-pull average consensus algorithm, based on graph theory and distributed optimization theory, abstracts various devices in the power system as intelligent agent nodes in the network, forming a communication topology graph through predefined neighbor relationships. Each agent node maintains local state information and control objectives, gradually achieving globally consistent control decisions through information exchange with neighbor nodes. Push operations enable agents to proactively propagate important state changes to neighbors, accelerating the spread of information throughout the network. Pull operations enable agents to proactively obtain the latest state information from neighbors, ensuring decisions are based on the most complete available information. Dynamic compression technology adaptively adjusts the accuracy and frequency of data transmission according to information importance and network conditions, significantly reducing the communication burden while ensuring control performance.

[0103] The detailed implementation steps include:

[0104] The first stage is network topology construction. The system defines neighbor relationships between agents based on the physical connections of power equipment and the communication network structure. The system uses a weighted undirected graph model to represent the network topology, with edge weights reflecting the strength of physical connections and communication quality between nodes. The system implements a topology adaptive adjustment mechanism; when a degradation in communication link quality or a node failure is detected, it automatically recalculates the optimal neighbor relationships to ensure network connectivity and efficient information propagation. The system also establishes a topology redundancy mechanism, configuring multiple communication paths for critical nodes to improve system robustness and reliability.

[0105] During the intelligent agent initialization phase, the system creates a corresponding intelligent agent for each power device, configuring an independent computing unit and communication module. The agent maintains local status including the device's current operating parameters, control objectives, constraints, and historical information. The system designs specialized agent templates for different types of equipment: energy storage system agents focus on state of charge and charge / discharge efficiency; controllable load agents focus on power levels and user comfort constraints; and distributed power generation agents focus on generation capacity and device health status. Each agent is equipped with a local decision-making module, enabling it to independently make basic control decisions during communication interruptions, ensuring continuous system operation.

[0106] During the push operation implementation phase, the system employs an efficient state push mechanism. Each agent proactively sends update information to its neighbors upon detecting significant state changes. The system uses an incremental update strategy, transmitting only the state change amount rather than the complete state vector, significantly reducing the amount of communication data. Push information includes the state vector, change gradient, and confidence index, helping the receiver assess the reliability and timeliness of the information. The system implements a priority push mechanism, giving the highest priority to critical state changes such as device failures and approaching safety limits, ensuring timely delivery of important information. The system also incorporates an adaptive push frequency mechanism, dynamically adjusting the push frequency based on the stability of the system's operating state. During stable periods, the push frequency is reduced to conserve communication resources, while during fluctuating periods, the push frequency is increased to ensure timely information delivery.

[0107] During the information retrieval phase, the system configures an intelligent information retrieval module for each agent, proactively obtaining the necessary state information from neighboring nodes. The retrieval operation employs a demand-based triggering mechanism; when an agent detects insufficient local information to make an optimal decision, it proactively sends information requests to relevant neighbors. The system implements a selective retrieval strategy, where agents determine the most needed information type based on current control objectives and constraints, requesting specific information selectively to avoid unnecessary communication overhead. After validity verification, the retrieved information enters the local fusion processing module. The system uses a weighted average method to integrate multi-source information, with weights dynamically adjusted based on the credibility and timeliness of the information sources.

[0108] In the adaptive quantization implementation phase, the system constructs a dynamic compression mechanism based on information theory principles, assessing the importance and uncertainty of each type of information according to Shannon entropy. The system establishes a three-level quantization system: key control signals use 16-bit high-precision quantization to ensure control accuracy; status monitoring information uses 8-bit medium-precision quantization to balance accuracy and bandwidth requirements; and auxiliary information uses 4-bit low-precision quantization to save communication resources. The system implements a context-aware quantization adjustment mechanism, monitoring the variance and rate of change of the system's operating state in real time. When the system is in a steady state, it automatically switches to coarse quantization mode to reduce communication overhead; when a system disturbance is detected, it immediately switches to fine quantization mode to ensure control accuracy.

[0109] During the consensus-building and execution phases, the system establishes multi-level convergence criteria to ensure reliable algorithm convergence. The first-level criterion monitors state changes in each agent based on state convergence; the second-level criterion evaluates the global optimization progress based on objective function convergence; and the third-level criterion measures the degree of coordination among agents based on consistency error. The system employs a hierarchical convergence strategy, allowing different control levels to converge at different speeds. Urgent control actions are prioritized for consensus and rapid execution, while non-critical adjustments allow for longer negotiation times to ensure optimality. When convergence conditions are met, all agents lock their current consensus state and synchronously execute corresponding power adjustment operations. The system ensures the coordination and consistency of agent actions through a time synchronization protocol.

[0110] Step 5: Closed-loop feedback and adaptive optimization;

[0111] Based on the actual execution status of the equipment and the operating status of the system, the hybrid prediction model and scheduling strategy are continuously optimized through reinforcement learning algorithms to form an adaptive closed-loop control system.

[0112] In step five, a real-time data feedback channel is established to collect actual execution status data through the SCADA system and smart meter acquisition devices, with a data update frequency of 1 minute. Prediction error indicators are calculated, including mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE). A model retraining mechanism is triggered when any indicator exceeds a preset threshold. An online learning algorithm is designed, using incremental learning to update the prediction model parameters. The weight of new samples is set to 0.1, and the weight of historical samples is adjusted according to an exponential decay law with a decay factor of 0.95. A scheduling performance evaluation system is constructed, including economic, safety, and environmental indicators. When the performance indicators deviate from the target value by more than 5%, an adaptive adjustment mechanism for the scheduling strategy is triggered. The Deep Deterministic Policy Gradient Algorithm (DDPG) is used to continuously optimize the scheduling strategy, taking the system state as input and the scheduling action as output. A multi-objective weighted sum is used as the reward function, and historical state-action-reward sequences are stored in an experience replay buffer to achieve continuous optimization of the scheduling strategy, forming a complete adaptive closed-loop control system.

[0113] In one embodiment, step one: multi-source data acquisition and intelligent preprocessing, specifically includes:

[0114] Step 1.1: Collect historical power generation data from wind power and photovoltaic power plants and real-time meteorological data from meteorological observation stations through the SCADA system, and perform timestamp consistency verification on the historical power generation data and real-time meteorological data to obtain synchronized raw data;

[0115] Specifically, historical power generation data from wind and solar power plants is collected through the SCADA system at a sampling frequency of 15 minutes. Data types include active power, reactive power, voltage, current, and other operating parameters. Simultaneously, real-time meteorological data from meteorological observation stations is collected, including wind speed, wind direction, temperature, humidity, air pressure, and solar irradiance, ensuring the consistency of data timestamps.

[0116] Step 1.2: Perform quality checks on the synchronized raw data to identify missing data, outliers, and duplicate records. Use linear interpolation to fill in missing values ​​and use box plots to identify outliers to generate cleaned data.

[0117] Linear interpolation for missing values ​​is a data repair technique based on the linear relationship between adjacent data points. This method assumes a linear relationship between the missing data point and its immediate neighbors, estimating the missing value by calculating the rate of change between adjacent points. Box plots are an outlier detection technique based on statistical distribution characteristics, identifying anomalous data points that deviate from the normal range by analyzing the quartile distribution of the data. Combining these two methods can effectively address common data quality issues in renewable energy generation data, providing a high-quality training dataset for subsequent predictive modeling.

[0118] Linear interpolation, based on the mathematical theory of linear approximation, assumes that the change in renewable energy power generation exhibits an approximately linear characteristic within a short time window. This method estimates the value of intermediate missing points by establishing a linear equation between adjacent valid data points before and after a missing point, utilizing the continuity and monotonicity of linear functions. Linear interpolation is particularly suitable for handling short time intervals in continuous time series, maintaining the temporal continuity and consistency of the data's trend. Box plots, based on descriptive statistics, characterize the data distribution by calculating the first quartile, third quartile, and interquartile range. This method utilizes the characteristics of normally or approximately normally distributed data to identify data points outside the reasonable statistical range as outliers, offering advantages such as simple calculation and wide applicability.

[0119] Step 1.2 begins with data preprocessing and missing value detection. The system performs a comprehensive scan of the raw data, establishing a time series integrity check mechanism. The system checks the data integrity at each time point according to a preset sampling interval, identifying the specific location and duration of missing data. A missing data pattern classification mechanism is established, categorizing missing data into three types: random missing, continuous missing, and periodic missing, and employing corresponding processing strategies for each type. The system also performs missing value cause analysis, examining equipment maintenance records, communication fault logs, and changes in meteorological conditions to determine possible causes of missing data, providing a reference for subsequent processing.

[0120] During the linear interpolation implementation phase, the system performs refined interpolation processing on the identified missing data points. For single-point missing data, the system directly uses linear interpolation of adjacent points to fill in the gaps, determining the rate of change by calculating the numerical and temporal differences between adjacent points, and then estimating the value of the missing point. For consecutive missing data with a short duration, the system adopts a piecewise linear interpolation strategy, dividing the entire missing interval into several sub-intervals, and using a linear change assumption for interpolation within each sub-interval. The system sets a time threshold for consecutive missing data; when the consecutive missing time exceeds two hours, linear interpolation may produce large errors, at which point the system automatically switches to a historical data replacement strategy. The system maintains a historical database storing data from the same time period over the past seven, fourteen, and twenty-one days, and calculates the replacement value using a weighted average method, with weights determined based on time distance and data similarity.

[0121] In the box plot anomaly detection implementation phase, the system calculates various statistical characteristic parameters based on the interpolated complete dataset. The system calculates the first quartile, third quartile, and interquartile range for each monitored parameter; these statistics objectively reflect the distribution characteristics and dispersion of the data. The system employs standard box plot anomaly detection criteria, marking data points exceeding the range of the first quartile minus 1.5 times the interquartile range or the third quartile plus 1.5 times the interquartile range as potential outliers. The system also implements a dynamic threshold adjustment mechanism, adjusting the stringency of anomaly detection based on the physical characteristics and historical distribution features of different parameters. For parameters with high inherent volatility, the detection threshold is appropriately relaxed, while for parameters with good stability, a stricter detection standard is applied.

[0122] In the outlier verification and handling phase, the system performs secondary verification on marked potential outliers, combining time continuity analysis and physical rationality checks for a comprehensive judgment. The system examines the data change trend before and after the outlier point; when the numerical change of the outlier point significantly deviates from the overall trend and the magnitude of the change exceeds the equipment's physical capabilities, it is confirmed as a true outlier. The system establishes an outlier classification mechanism, categorizing outliers into different types such as measurement errors, equipment failures, extreme operating conditions, and data transmission errors, and employs corresponding handling methods for each type. For clear measurement and transmission errors, the system directly deletes the outlier data and fills it in using interpolation methods; for outliers caused by equipment failures, the system retains the data but marks it specially, giving it lower weight during modeling; for outlier data under extreme operating conditions, the system retains it but adds an uncertainty indicator.

[0123] During the data integrity verification and quality assessment phase, the system conducts a comprehensive quality check and consistency verification on the cleaned data. The system calculates quality indicators such as data integrity rate, outlier rate, and interpolation rate to evaluate the effectiveness and reliability of the data cleaning. The system performs cross-parameter consistency checks to verify whether related parameters meet basic physical relationships, such as the correlation between wind speed and wind power, or solar irradiance and photovoltaic power. The system establishes a data quality scoring mechanism, comprehensively considering multiple dimensions such as completeness, accuracy, consistency, and timeliness, assigning a quality score to each data segment to provide a data credibility reference for subsequent modeling. The system also generates a detailed data cleaning report, recording the specific methods for filling missing values, the number and type of outliers identified, and the detailed data correction process, ensuring the traceability and transparency of the data processing.

[0124] Specifically, step 1.2 includes:

[0125] Step 1.2.1: Perform missing value detection on the synchronized raw data. When the continuous missing time is less than 2 hours, use linear interpolation to calculate the missing value based on the data before and after the time. When the continuous missing time exceeds 2 hours, use historical data from the same period to replace it and generate the completed data.

[0126] In the missing value detection and processing stage, the system first performs a comprehensive scan of the synchronized raw data to identify missing points in the time series. When a continuous missing period of less than 2 hours is detected, linear interpolation is used for imputation. This involves calculating the slope based on the valid data before and after the missing point and using this to estimate the value of the missing period. Specifically, for a missing value at time t, the system finds the nearest valid data points t1 and t2 before and after it, and calculates the interpolation using the formula v(t) = v(t1) + (v(t2) - v(t1)) × (t - t1) / (t2 - t1). When the continuous missing period exceeds 2 hours, linear interpolation may lead to significant errors. In this case, the system automatically switches to a historical data replacement strategy, finding the average value of data from the same period over the past 7, 14, or 21 days as the replacement value to ensure the reasonableness of the imputed data. The system also adjusts the imputation strategy according to the characteristics of the data type. For example, for meteorological data such as wind speed and light intensity, its periodicity is considered; for equipment status information, the nearest valid value imputation strategy is used. Through these targeted processing steps, the system generates a completed dataset that is time-continuous and numerically reasonable.

[0127] Step 1.2.2: Based on the completed data, calculate the first quartile, third quartile, and interquartile range of each parameter, identify outlier data points exceeding 1.5 times the interquartile range using the box plot method, and generate outlier labeling results;

[0128] In the outlier identification and processing stage, the system first calculates the statistical characteristics of each parameter based on the completed dataset. For each parameter sequence, it calculates the first quartile Q1 (the value at the 25th percentile after sorting), the third quartile Q3 (the value at the 75th percentile after sorting), and the interquartile range (IQR) (Q3-Q1). The system uses the classic box plot method to define outlier boundaries: the lower boundary is Q1-1.5×IQR, and the upper boundary is Q3+1.5×IQR. Any data point outside this range is marked as a potential outlier. To improve the accuracy of anomaly detection, the system also performs a secondary judgment based on the temporal continuity of the data. When a point is marked as an anomaly, the system examines the trend of changes in the data points before and after it. If the change is too drastic and does not conform to physical laws, it is confirmed as an anomaly; otherwise, it may be normal fluctuation and is retained. The system performs anomaly detection independently for each parameter, generating a complete result set containing anomaly labels, providing a basis for subsequent data cleaning.

[0129] Step 1.2.3: Remove abnormal data points according to the outlier marking results, and perform integrity verification on the removed data to ensure data continuity and consistency, thereby obtaining the cleaned data.

[0130] In the data cleaning and integrity verification phase, the system systematically removes data points identified as anomalous based on the aforementioned outlier labeling results. For isolated outliers, the same imputation strategy as for missing values ​​is directly applied; for continuous outlier segments, an appropriate processing method is selected based on their length. After outlier handling, the system performs rigorous data integrity verification, including: time series continuity checks to ensure uninterrupted data flow over time; numerical range reasonableness checks to verify that all parameter values ​​are within physically feasible ranges; and correlation consistency checks to verify that interrelated parameters (such as wind speed and wind power) satisfy basic physical relationships. The system also calculates key statistical indicators such as mean, standard deviation, skewness, and kurtosis, comparing them with historical data to ensure that the cleaned dataset retains the statistical characteristics of the original data. Through this series of rigorous verifications, a high-quality, continuous, and consistent cleaned dataset is finally obtained, providing a reliable foundation for subsequent model training.

[0131] Step 1.3: Perform Min-Max normalization on the cleaned data to map the data to the interval between 0 and 1, extract the periodic features of hours, days of the week, and months, and generate lag features to obtain the normalized training dataset.

[0132] Specifically, all numerical data are subjected to Min-Max normalization, mapping the data to the [0,1] interval to eliminate the influence of different units on model training. A time feature engineering process is implemented to extract periodic features such as hours, days of the week, and months, and lag features are generated to enhance time series modeling capabilities.

[0133] Step 2: Construction of a neural symbol fusion-based power generation prediction model, including:

[0134] Step 2.1: Based on the normalized training dataset, the power change rate constraint and device capacity constraint are encoded into symbolic rules, and the constraint rules are embedded into the deep learning model through the neural symbolic interface layer;

[0135] Specifically, a neural symbolic architecture with time logic constraints is designed to encode physical constraints (such as power change rate limits and device capacity constraints) into symbolic rules, and these constraints are embedded into the deep learning model through a neural symbolic interface layer.

[0136] In the neural symbolic interface layer construction phase, the system, based on a normalized training dataset, first explicitly encodes the physical constraints of the renewable energy power generation system into symbolic rules. These constraints mainly include power rate of change limits (characterizing equipment ramp-up capabilities), equipment capacity constraints (ensuring predicted power does not exceed installed capacity), minimum output constraints (considering equipment start-up and shutdown characteristics), and environmental condition constraints (such as wind speed cut-in and cut-out characteristics). The system uses first-order predicate logic to represent these constraints; for example, the power rate of change constraint is represented as follows:

[0137]

[0138] Where δ is the maximum allowable rate of change coefficient. These symbolic rules are connected to the deep learning model through a specially designed neural symbolic interface layer. The interface layer adopts a differentiable soft constraint implementation method, which transforms the hard constraints into penalty terms of the neural network, enabling the model to gradually learn to satisfy these physical constraints during training, while maintaining end-to-end differentiability and ensuring that gradients can be effectively propagated.

[0139] Step 2.2: Construct a time logic constraint processing module to encode the constraint that the wind power change rate does not exceed 10% of the rated power per minute and the photovoltaic power change rate does not exceed 20% of the rated power per minute, and generate constraint embedding vectors;

[0140] Specifically, time logic constraints are introduced to ensure that the predicted power change rate does not exceed the equipment ramp-up capability (10% / minute of rated power for wind power and 20% / minute for photovoltaic power).

[0141] In the construction of the temporal logic constraint processing module, the system designed dedicated temporal logic constraint processing mechanisms for the different physical characteristics of wind power and photovoltaic power generation. For wind power systems, considering their mechanical inertia, the system imposes strict constraints on the rate of power change, limiting it to no more than 10% of the rated power per minute. This constraint is achieved through a temporal convolutional layer followed by a constraint encoder, which converts the physical constraints into high-dimensional embedding vectors as modulation signals for subsequent network layers. For photovoltaic systems, considering their faster response speed, the system sets a relatively relaxed rate of change constraint of 20% of the rated power per minute, while also adding physical correlation constraints between light intensity and power output. The system also considers special processing for sunrise and sunset periods, using solar altitude angle information provided by the astronomical calculation module to apply special constraints to the prediction results for these periods. These temporal logic constraints are encoded into structured embedding vectors, providing constraint information at different levels of the model to ensure that the prediction results conform to physical laws.

[0142] Step 2.3: Capture short-term temporal dependencies using an LSTM network;

[0143] In the short-term prediction stage of the LSTM network, the system constructs a multi-layer stacked LSTM network architecture to capture short-term temporal dependencies. The network's input features include historical data from the past 24 hours, specifically: power generation time-series data (sampling interval 15 minutes), meteorological data (wind speed, wind direction, temperature, humidity, light intensity, etc.), and equipment status information (such as equipment availability, maintenance records, etc.). The LSTM network adopts a 3-layer stacked structure, with each layer having a hidden state dimension of 128. Forget gates, input gates, and output gates effectively capture long-term and short-term dependencies in the time-series data. To enhance the model's ability to perceive key time points, the system introduces an attention mechanism, adaptively selecting key information by calculating importance weights at different time steps. The attention weights are calculated using softmax(tanh(W·h)). t Calculate + b)), where h t This is a hidden state. The model output is the hourly power prediction value for the next 6 hours, and the post-processing module ensures that the prediction result meets the aforementioned time logic constraints, especially the power change rate does not exceed the limit of the equipment's ramp-up capability, so as to obtain a physically reasonable short-term power prediction result.

[0144] Specifically, step 2.3 includes:

[0145] Step 2.3.1: Based on the normalized training dataset, construct a multi-layer LSTM network with a hidden layer dimension of 128, and adaptively select information from key time steps through an attention mechanism;

[0146] In the LSTM network construction phase, the system meticulously designed a multi-layer LSTM network architecture based on a normalized training dataset to effectively capture short-term temporal dependencies. This network employs a three-layer stacked structure, with each layer's hidden state dimension set to 128. This dimension was optimized through grid search, achieving a good balance between model complexity and computational efficiency. To enhance the network's ability to perceive key time points, the system introduces an attention layer based on the Bahdanau mechanism, calculating attention weights α. t = softmax(v T ·tanh(W·h t + U·s t-1 + b)), where h t For the encoder's hidden state, s t-1The decoder uses the previous state to adaptively select key time steps from historical data. The attention mechanism significantly improves the model's predictive ability for abrupt changes and turning points, especially during periods of drastic weather change. The network's input layer employs a feature fusion design, processing time features (hour, date, and season encoding), historical power data, and meteorological features through independent embedding layers and then concatenating them to form a rich input representation. To prevent overfitting, the system inserts Dropout layers (with a dropout rate of 0.2) and BatchNormalization layers between the LSTM layers, significantly improving the model's generalization ability and training stability.

[0147] Step 2.3.2: Input the power generation, meteorological data and equipment status information of the past 24 hours into the LSTM network for processing, extract short-term time-series dependency features, and generate hourly power prediction values ​​for the next six hours;

[0148] In the LSTM network processing and feature extraction stage, the system performs deep processing on multi-source data from the past 24 hours (i.e., 96 15-minute sampling points). Input data includes historical power generation sequences, meteorological data (wind speed, wind direction, temperature, humidity, air pressure, light intensity, etc.), and equipment status information (such as operating mode, available capacity, etc.). The system first performs time alignment and feature normalization on this heterogeneous data, and then constructs input-output sample pairs using a sliding window technique. The LSTM network effectively manages long short-term memory through a gating mechanism, especially the forgetting gate. t = σ(W f ·[h t-1 , x t ] + b f To control the degree of retention of historical information, input gate i t = σ(W i ·[h t-1 , x t ] + b i The output gate determines the degree to which new information is accepted. t = σ(W o ·[h t-1 , x t ] + b o The system controls the output of information. This mechanism enables the network to adaptively capture key patterns and dependencies in time-series data, particularly for predicting power change trends. Through a bidirectional LSTM structure, the system can simultaneously consider past and future contextual information, further enhancing the comprehensiveness of feature extraction. The processed features are mapped to the output space through fully connected layers to generate hourly power predictions for the next six hours, with a prediction resolution of 15 minutes, meeting the refined requirements of power dispatching.

[0149] Step 2.3.3: Introduce time logic constraints to ensure that the predicted power change rate does not exceed the equipment ramp-up capability, and obtain short-term power prediction results that meet physical constraints.

[0150] Temporal logic constraints are a technique that embeds the dynamic characteristics and operational limitations of a physical system into a predictive model. This constraint ensures that the prediction results conform to the actual physical capabilities and operational patterns of the equipment. Equipment ramp-up capability refers to the maximum magnitude by which a power generation device can change its output per unit time; this physical characteristic directly determines the speed and scope of the device's response to external dispatch commands. By introducing temporal logic constraints into the predictive model, the system can generate prediction results that accurately reflect future trends while strictly satisfying physical feasibility, providing reliable technical support for power system dispatch decisions.

[0151] Time-logic constraints are based on the state transition theory of physical systems and reachability analysis in cybernetics. This principle posits that the state changes of any physical system are limited by its inherent characteristics, preventing infinitely large instantaneous jumps. For renewable energy power generation equipment, these limitations primarily manifest in the mechanical inertia, electrical characteristics, and response speed of the control system. Wind turbines are constrained by the aerodynamic characteristics of the blades, the transmission characteristics of the gearbox, and the electromagnetic characteristics of the generator, resulting in a definite physical upper limit to their power change rate. While photovoltaic power generation systems have relatively fast response speeds, they are still limited by inverter switching speed, maximum power point tracking algorithm response time, and grid synchronization requirements. Time-logic constraints express these physical limitations as computable constraints through mathematical modeling, ensuring that the output of the prediction model always remains within a physically feasible range.

[0152] Specifically, this step begins with the parameterized modeling stage of equipment ramp-up capability. The system establishes detailed ramp-up capability models for different types of renewable energy equipment. For wind turbines, based on the technical specifications provided by the equipment manufacturer and actual operating data, the system determines the maximum power change rate under rated power conditions, typically set at 10% of the rated power per minute. The system also considers the impact of environmental conditions on ramp-up capability, establishing a correlation model between environmental factors such as temperature, humidity, and altitude and ramp-up capability, dynamically adjusting ramp-up limits under different environmental conditions. For photovoltaic power generation systems, the system sets a relatively high power change rate limit, typically 20% of the rated power per minute, while also considering the impact of factors such as cloud cover and changes in sunlight angle on the rate of power change. The system establishes a model of the impact of equipment aging and maintenance status on ramp-up capability, appropriately adjusting the ramp-up capability parameters as equipment operating time increases and maintenance status changes.

[0153] In the constraint mathematical modeling phase, the system transforms physical ramp limits into mathematical constraint expressions, establishing a model relating the power change rate to time intervals. The system defines power change constraints between consecutive time points to ensure that the power difference between adjacent prediction points does not exceed the maximum possible change in power within the corresponding time interval. The system establishes a multi-timescale constraint system, considering both short-term (minute-level) and medium-term (hour-level) cumulative change constraints to prevent situations where individual changes meet requirements but cumulative changes exceed capacity. The system also introduces directional constraints, considering potential asymmetric characteristics of the equipment in power increase and decrease directions, and sets ramp limits for both upward and downward adjustments.

[0154] In the constraint embedding neural network design phase, the system employs a combination of soft and hard constraints to integrate temporal logical constraints into the prediction model. For soft constraints, the system adds a constraint penalty term to the loss function. When a prediction violates the climbing constraint, additional penalties are imposed, guiding the model to learn prediction patterns that satisfy the constraints. The weights of the penalty term use an adaptive adjustment strategy: smaller weights are used in the early stages of training to allow the model to fully learn data patterns, while the weights are gradually increased in the later stages of training to emphasize the importance of the constraints. For hard constraints, the system adds a constraint projection module after the model's output layer. An iterative projection algorithm projects the constraint-violation prediction sequences into the feasible region, ensuring that the final output strictly satisfies all constraints.

[0155] In the constraint projection algorithm implementation phase, the system designs an efficient constraint projection algorithm to transform the original prediction sequence that does not meet the constraints into a feasible prediction sequence that does meet the constraints. The system employs a sequential dynamic programming method, starting from the starting point of the prediction sequence and progressively checking and adjusting the value of each prediction point to ensure that its change from the previous point does not exceed the ramp-up limit. When a prediction point that violates the constraints is detected, the system calculates the optimal adjustment value for that point within the constraint range, while considering the impact on subsequent prediction points, and uses a global optimization method to determine the best adjustment strategy. The system also implements a constraint relaxation mechanism, allowing for appropriate relaxation of constraints within a safe range when strict ramp-up constraints cause a significant decrease in prediction accuracy, balancing the requirements of prediction accuracy and physical feasibility.

[0156] In the multi-constraint coordination and optimization phase, the system addresses the coordination problem between time logic constraints and other constraints, establishing a unified constraint optimization framework. Besides ramp-up constraints, the system also needs to simultaneously satisfy multiple restrictions such as power upper and lower limits, equipment availability constraints, and environmental condition constraints. The system establishes a constraint priority mechanism, setting safety-related constraints as the highest priority, performance-related constraints as medium priority, and economic-related constraints as lower priority, prioritizing constraints in case of conflict. The system employs a multi-objective optimization method to handle the trade-offs between constraints, seeking a prediction scheme that optimizes other objectives while satisfying key constraints.

[0157] In the constraint verification and result evaluation phase, the system comprehensively verifies and evaluates the prediction results after applying time logic constraints. The system checks whether all prediction points strictly meet the ramp-up constraints, statistically analyzes the number and severity of constraint violations, and assesses the effectiveness of constraint implementation. The system compares and analyzes the changes in prediction accuracy before and after constraints, quantifies the impact of physical constraints on prediction performance, and ensures that prediction accuracy is maintained as much as possible while satisfying physical feasibility. The system establishes an executability evaluation mechanism for prediction results, simulates the actual scheduling execution process, and verifies the operability and security of the prediction results in actual operation. The system also generates detailed constraint execution reports, recording the specific process of constraint adjustment, the statistics of constraint satisfaction, and a comprehensive evaluation of prediction performance, providing a scientific basis for model optimization and parameter adjustment.

[0158] For example, in the time logic constraint introduction stage, the system explicitly incorporates the ramp-up capability limit of the physical system into the prediction model to ensure that the prediction results meet the operational constraints of the actual equipment. For wind power systems, based on the mechanical characteristics of the turbine, the system sets the upper limit of the power change rate to 10% of the rated power per minute; for photovoltaic systems, considering their faster response speed, the upper limit of the change rate is set to 20% of the rated power per minute. These constraints are implemented in two ways: one is by adding a regularization term to the loss function to penalize predicted values ​​that violate the constraints, i.e.:

[0159] L total = L prediction + λ·L constraint ,in:

[0160] L constraint = max(0, |P(t) - P(t-1)| - δ×P rated ) 2 First, λ is a trade-off parameter. Second, a constraint projection layer is added after the model output layer. An iterative projection algorithm projects the prediction sequences that do not meet the constraints to the feasible solution space. The system also considers various physical constraints such as equipment start-up and shutdown characteristics and environmental conditions (such as cut-in and cut-out wind speeds). By combining soft and hard constraints, the system ensures that the final short-term power prediction results accurately reflect meteorological change trends and strictly meet the operational constraints of the physical system, providing a reliable foundation for subsequent scheduling optimization.

[0161] Step 2.4: Extract medium- to long-term spatial-temporal features using a CNN-Transformer network;

[0162] In the long-term prediction stage of the CNN-Transformer, the system employs an innovative hybrid architecture, combining the local feature extraction capabilities of CNNs with the long-range dependency modeling capabilities of Transformers. The CNN module uses a multi-scale one-dimensional convolutional structure, containing three sets of parallel convolutional layers with kernel sizes of 3, 5, and 7, each containing 64 convolutional filters. Different scale kernels capture temporal patterns across different periods. The convolutional outputs pass through a feature fusion layer, integrating multi-scale features into a unified representation. The Transformer module uses a standard encoder structure, incorporating a multi-head self-attention mechanism and a feedforward neural network. The self-attention mechanism uses eight attention heads, each with a dimension of 64, enhancing the model's expressive power by parallel computation of attention distributions across different representation subspaces. The feedforward network uses a two-layer structure with a hidden layer dimension of 512, employing the GELU activation function. To handle the positional information of temporal data, the system uses sinusoidal positional encoding to ensure the model can perceive the temporal order of data points. Through this hybrid architecture, the system can simultaneously capture local temporal patterns and long-term dependencies, providing robust support for medium- and long-term predictions.

[0163] Specifically, step 2.4 includes:

[0164] Step 2.4.1: Based on the normalized training dataset, design a CNN-Transformer hybrid architecture. The CNN module uses one-dimensional convolution to extract local temporal patterns, and the convolution kernel sizes are 3, 5, and 7 respectively.

[0165] In the CNN-Transformer hybrid architecture design phase, the system constructs an innovative deep learning architecture based on a normalized training dataset, effectively integrating the local feature extraction capabilities of CNNs and the long-range dependency modeling capabilities of Transformers. The CNN module employs a multi-scale parallel design, containing three independent one-dimensional convolutional layers with kernel sizes of 3, 5, and 7, corresponding to receptive fields covering time ranges of 45 minutes, 75 minutes, and 105 minutes, respectively. Each convolutional layer contains 64 filters, capturing local patterns across different time spans through kernels of varying scales. The convolutional layers utilize causal convolution to ensure the model does not exploit future information, meeting the causal requirements of time-series prediction. The convolutional outputs are processed through a LeakyReLU activation function (with a negative slope of 0.1) and residual connections, effectively mitigating the vanishing gradient problem and accelerating training convergence. Multi-scale features are integrated through an attention fusion mechanism; the system calculates the importance weights of features at each scale, achieving adaptive feature selection and significantly enhancing the model's ability to capture multi-period patterns, particularly for comprehensive modeling of intraday, intraweekly, and seasonal variations.

[0166] Step 2.4.2: Capture the changing patterns of different cycles through multi-scale feature fusion. The Transformer module uses a multi-head self-attention mechanism to handle long-term dependencies, with the number of heads set to 8 and the feedforward network dimension set to 512.

[0167] In the Transformer module implementation, the system adopts a standard encoder-decoder architecture, but with special optimizations for temporal prediction tasks. The encoder contains 6 self-attention layers, each consisting of a multi-head self-attention layer and a feedforward neural network. The multi-head self-attention mechanism uses 8 attention heads, each with a dimension of 64, expressed by the formula Attention(Q,K,V) = softmax(QK). T / d k The attention distribution is calculated using V to achieve parallel attention to information at different locations. The feedforward network adopts a two-layer structure with a hidden layer dimension of 512, using the GELU activation function, which has smoother gradient characteristics compared to ReLU, thus benefiting model training. To handle the location information of temporal data, the system employs a sinusoidal position encoding method, using PE(pos,2i) = sin(pos / 10000) 2i / d_model ) and PE(pos,2i+1) = cos(pos / 10000) (2i / d_model) The system computes positional encoding to ensure the model can perceive the temporal order of data points. The decoder also contains six layers, but incorporates a cross-attention mechanism, enabling the model to focus on key information in the encoder output. To enhance the model's ability to process long sequences, the system also introduces relative positional encoding and local attention mechanisms, effectively expanding the model's effective receptive field and improving its ability to model long-term dependencies.

[0168] Step 2.4.3: The prediction problem is modeled as a constrained optimization problem using neural symbolic acceleration technology. The approximate optimal solution is quickly solved by neural network, and the result is verified and corrected by symbolic reasoning to obtain the medium- and long-term space-time characteristics. The medium- and long-term space-time characteristics are the medium- and long-term power prediction results that meet the physical feasibility constraints.

[0169] The neural symbolic acceleration technology is one of the innovative aspects of this application's embodiments. This technology is an advanced artificial intelligence method that integrates neural networks and symbolic reasoning. It combines the powerful data processing capabilities of deep learning with the logical reasoning capabilities of symbolic systems, achieving an organic unity between data-driven modeling and knowledge-driven reasoning. This technology transforms complex prediction problems into constrained optimization problems, utilizes the rapid approximation capabilities of neural networks to generate preliminary solutions, and then ensures the correctness and feasibility of the solutions through logical verification and knowledge correction by the symbolic reasoning system. This hybrid approach maintains the advantages of neural networks in handling complex nonlinear relationships while fully leveraging the strengths of symbolic reasoning in ensuring logical consistency and interpretability.

[0170] The neurosymbolic acceleration technology, based on a theoretical framework that integrates computational intelligence and cognitive intelligence, decomposes the prediction task into two complementary stages: rapid approximation and precise reasoning. The neural network component employs an end-to-end learning approach, training on a large amount of historical data to obtain a complex mapping relationship from input features to prediction results, enabling it to quickly generate near-optimal predictions. The symbolic reasoning component constructs a rule base based on domain knowledge and physical laws, verifying the rationality of the neural network output through logical reasoning and identifying and correcting potentially inconsistent or infeasible predictions. The two components exchange information through a carefully designed interface layer: the neural network provides the numerical computation foundation for symbolic reasoning, while symbolic reasoning provides logical constraints and knowledge guidance to the neural network. This collaborative mechanism allows the system to handle complex data patterns while ensuring the physical rationality and logical consistency of the results.

[0171] In S2.4.3, the first stage is the constrained optimization problem modeling phase, where the system transforms the medium- to long-term power forecasting task into a standard mathematical optimization problem. The system defines the decision variables as the predicted power values ​​at each time point within the next week. The objective function is designed to minimize the expected deviation between the predicted and actual observed values, while considering penalties for forecast smoothness and uncertainty. The system establishes a complete set of constraints, including upper and lower power limits to ensure the predicted values ​​remain within equipment capacity, ramp-up constraints to ensure power changes between adjacent time points do not exceed equipment capacity, continuity constraints to ensure temporal consistency of the forecast sequence, and environmental constraints to ensure a reasonable correspondence between the predicted values ​​and meteorological conditions. The system also introduces uncertainty constraints, requiring the confidence interval of the forecast results to cover specific percentiles of the historical error distribution to ensure the statistical reliability of the forecast.

[0172] In the rapid solution stage of the neural network, the system employs deep learning methods to construct an efficient approximate solver capable of generating near-optimal solutions to optimization problems within seconds. The system utilizes a dedicated neural network architecture. The input layer receives weather forecast data, historical power data, equipment status information, and the parameter settings of the optimization problem, extracting complex feature relationships through multi-layer nonlinear transformations. The intermediate layers employ attention mechanisms and residual connections to enhance the model's ability to perceive key information and improve gradient propagation efficiency. The output layer directly generates the predicted power sequence, along with confidence assessment and uncertainty quantification information. The system employs a multi-task learning framework, simultaneously optimizing prediction accuracy, constraint satisfaction, and computational efficiency, balancing the relationships between different objectives through dynamic weight adjustments. Network training combines reinforcement learning and supervised learning. The supervised learning part uses historical data to learn basic prediction patterns, while the reinforcement learning part learns better solution strategies through interaction with the optimization solver.

[0173] In the symbolic reasoning verification phase, the system constructs a domain-knowledge-based symbolic reasoning engine to perform comprehensive logical verification and consistency checks on the prediction results generated by the neural network. The system establishes a hierarchical rule base, including basic rules at the physical law level such as energy conservation and power balance, technical rules at the equipment characteristic level such as ramp limits and efficiency curves, and heuristic rules at the operational experience level such as seasonal patterns and abnormal operating condition identification. The reasoning engine employs a strategy combining forward and backward reasoning. Forward reasoning derives possible conclusions from known conditions, while backward reasoning searches for supporting evidence and logical chains from the target conclusion. The system implements a multi-layered verification mechanism: first, a basic numerical range check is performed to ensure the predicted value is within a reasonable range; then, a physical consistency check verifies the logical relationship between the predicted value and the input conditions; and finally, a global consistency check ensures the internal logical consistency of the entire prediction sequence.

[0174] During the result correction and optimization phase, the system intelligently corrects and locally optimizes the initial predictions of the neural network based on the verification results of symbolic inference. When the inference engine identifies inconsistencies or infeasible parts in the prediction results, the system activates an automatic correction mechanism, employing appropriate correction strategies based on the type and severity of constraint violations. For minor constraint violations, the system uses a local adjustment method, fine-tuning specific time points while maintaining the overall prediction pattern. For severe logical conflicts, the system uses a global reconstruction method, reconstructing some prediction sequences based on the logical framework provided by symbolic inference. The correction process employs an iterative optimization strategy, re-verifying the symbol after each correction until all constraints are satisfied or a preset maximum number of iterations is reached. The system also implements a correction effect evaluation mechanism, comparing changes in prediction quality before and after correction to ensure that the correction process improves rather than deteriorates prediction performance.

[0175] In the medium-to-long-term spatial-temporal feature extraction phase, the system extracts physically meaningful spatial-temporal features from the validated and corrected prediction results, providing structured input information for subsequent scheduling optimization. Spatial feature extraction focuses on the power correlation and complementarity between different geographical locations. The system analyzes the output coordination relationships among multiple wind farms or photovoltaic power plants and identifies the impact of regional weather patterns on overall power generation capacity. Temporal feature extraction focuses on power variation patterns at different time scales. The system identifies rapid fluctuations at the hourly level, periodic changes at the daily level, meteorological pattern transitions at the weekly level, and long-term trends at the seasonal level. The system employs mathematical tools such as principal component analysis and independent component analysis to extract low-dimensional key features from high-dimensional prediction data, preserving important variation information while reducing data processing complexity. The extracted features undergo standardization and normalization to ensure that features of different dimensions and scales can be compared and integrated within a unified framework.

[0176] During the quality assurance and performance evaluation phase, the system establishes a comprehensive quality control system to fully evaluate the execution effect and prediction quality of the neural symbolic acceleration technology. The system designs multi-dimensional evaluation metrics, including prediction accuracy metrics such as mean absolute error and root mean square error, constraint satisfaction metrics such as the number and severity of constraint violations, computational efficiency metrics such as solution time and resource consumption, and interpretability metrics such as the clarity and logical consistency of the inference path. The system establishes a benchmark comparison mechanism to compare the results of the neural symbolic acceleration technology with traditional pure neural network methods, pure optimization methods, and expert experience methods, quantifying the performance improvement and advantages of the new technology. The system also implements an online performance monitoring mechanism to track changes in key metrics during the technology's execution process in real time, promptly identifying and handling potential performance degradation or anomalies to ensure the stability and reliability of the technology in practical applications.

[0177] Specifically, in the application of neuro-symbolic acceleration technology, the system models the prediction problem as a constrained optimization problem. Through the collaborative work of neural networks and symbolic reasoning, it achieves high-precision and physically feasible prediction results. First, the system uses the aforementioned CNN-Transformer network to generate a preliminary prediction result P. initial This serves as the starting point for the optimization problem; then, the objective function is constructed as min P ||P - P initial || 2The system imposes multiple constraints, including power upper and lower bound constraints (0 ≤ P(t) ≤ Pmax(t), ramp constraints (|P(t) - P(t-1)| ≤ δ×Prated), and environmental constraints (such as the nonlinear relationship between wind speed and power). A highly efficient mixed-integer linear programming (MILP) solver is employed to handle this optimization problem. Piecewise linearization is used to address the nonlinear constraints, significantly improving solution efficiency. To further accelerate the computation, a neural-guided heuristic search strategy is implemented, utilizing the confidence intervals predicted by the neural network to guide the optimization search direction and drastically reduce the search space. Finally, the system verifies and corrects the optimization results using a symbolic inference engine, checking for predictions that violate physical laws and adjusting them according to domain knowledge rules. This neural-symbolic fusion approach fully combines the flexibility of data-driven models with the interpretability of physical models. The generated medium- and long-term power prediction results not only have high accuracy but also strictly satisfy various constraints of the physical system, providing a reliable decision-making basis for power system dispatch.

[0178] Step 2.5: The prediction problem is modeled as a constrained optimization problem using the neural symbol acceleration technology of MILP motion planning. The approximate optimal solution is quickly solved by the neural network, and the prediction results are verified and corrected by symbolic reasoning to obtain the hybrid prediction model that satisfies the time logic constraints.

[0179] The Neural Symbolic Acceleration Technique for MILP Motion Programming is an advanced intelligent optimization method that combines mixed-integer linear programming theory, motion planning algorithms, and neural symbolic computation. This technique is specifically designed to handle dynamic system prediction problems with complex temporal logical constraints. It treats the renewable energy power generation prediction problem as a time-domain motion planning problem, where the evolution of the system state along the time axis must satisfy multiple constraints, including physical laws, equipment characteristics, and operational constraints. By introducing mixed-integer variables to represent discrete decision choices and continuous state changes, the system can accurately model complex temporal logical relationships. Simultaneously, by leveraging the rapid approximation capability of neural networks and the logical verification capability of symbolic reasoning, it achieves efficient and accurate constraint satisfaction prediction.

[0180] MILP motion planning theory, based on the mathematical framework of optimal control theory and path planning, transforms the time series forecasting problem into finding the optimal trajectory in the state space. The system treats the predicted power value at each time point as a state point in the state space, and the entire forecast sequence constitutes a trajectory from the initial state to the target state. Temporal logical constraints are modeled as feasible regions and transition constraints in the state space; the system must find the optimal predicted trajectory while satisfying these constraints. The mixed-integer characteristic is reflected in the system's need to simultaneously handle continuous power value changes and discrete operating mode selections, such as equipment start / stop states, operating mode switching, and maintenance schedule arrangements. The core of neural symbolic acceleration lies in using deep learning models to quickly generate initial trajectories that satisfy most constraints, and then using a symbolic inference system to perform precise logical verification and local corrections to the trajectory, ensuring both solution efficiency and the correctness and completeness of the solution.

[0181] In this embodiment of the application, step 2.5 is divided into several stages:

[0182] In the state-space modeling and trajectory planning framework construction phase, the system establishes a complete state-space representation to describe the dynamic characteristics of the renewable energy power generation system. The state vector contains multi-dimensional information such as the current power output value, equipment operating status, environmental condition parameters, and cumulative constraint states. These state variables comprehensively characterize the system's operating status at a specific moment. The system defines a state transition function to describe the system's evolution from one moment to the next, considering the physical characteristics of the equipment, the impact of environmental changes, and the effects of control inputs. The control input space includes all adjustment methods that the system can apply, such as adjusting the power setpoint, selecting the operating mode, and executing maintenance operations. The system establishes a trajectory evaluation function to measure the quality of different predicted trajectories. The evaluation function comprehensively considers multiple performance indicators such as prediction accuracy, trajectory smoothness, energy efficiency, and safety margin.

[0183] In the MILP modeling phase for time-logic constraints, the system transforms complex time-logic constraints into standard mixed-integer linear programming constraints. The system employs a linear time logic mathematical framework to express time-related constraints, such as complex logical relationships like "power output must remain within a specific range for the next three hours" or "the device must run continuously for at least two hours after startup before being shut down." The system introduces auxiliary binary variables to represent the truth or falsity of logical propositions and transforms logical constraints into a system of linear inequalities through linearization techniques of logical operations. The system handles time window constraints, establishing a sliding time window mathematical model to ensure that constraints are satisfied within any time window across the entire prediction time domain. The system also establishes a constraint priority mechanism, assigning different weights and priorities to constraints of varying importance, and enabling the selection and balancing of constraints according to predetermined rules when conflicts arise.

[0184] In the rapid trajectory generation stage, the system employs a dedicated deep neural network architecture to quickly generate initial predicted trajectories that satisfy most constraints. The network architecture adopts a sequence-to-sequence design pattern, with the encoder processing historical data and environmental information, and the decoder generating future predicted sequences. The system embeds a constraint-aware mechanism into the network, using an attention mechanism to focus on constraint-related features. Constraint embedding layers encode constraints into high-dimensional feature vectors and integrate them into the network's computation. Network training utilizes a multi-task learning strategy, simultaneously optimizing both prediction accuracy and constraint satisfaction, balancing these two objectives through a dynamic weight adjustment mechanism. The system implements a constraint-guided sampling mechanism, prioritizing the sampling of difficult samples near constraint boundaries during training to improve the network's learning performance on constraints. The network also integrates an uncertainty quantification module, providing confidence estimates for each prediction point to guide subsequent symbolic inference verification.

[0185] In the symbolic reasoning verification and correction phase, the system constructs a verification engine based on logical reasoning to rigorously check the constraints and perform logical verification on the initial trajectory generated by the neural network. The verification engine employs a hierarchical verification strategy: first, it performs basic numerical constraint checks to verify whether the predicted values ​​are within the allowable range; then, it performs temporal logical constraint checks to verify whether the trajectory satisfies complex time-related logical relationships; and finally, it performs a global consistency check to ensure the logical self-consistency of the entire trajectory. The system establishes a constraint violation diagnosis mechanism, which can accurately locate the specific constraint conditions, the time interval of the violation, and the severity of the violation when a constraint violation is detected. Based on the diagnostic results, the system initiates an intelligent correction algorithm, using local search and constraint propagation techniques to accurately correct the trajectory segments that violate the constraints. The correction process adopts the principle of minimum interference, preserving the shape and characteristics of the original trajectory as much as possible while satisfying the constraints, avoiding performance loss due to over-correction.

[0186] In the hybrid prediction model integration and optimization phase, the system organically combines three components: rapid neural network generation, symbolic inference verification and correction, and MILP motion planning optimization, to construct a complete hybrid prediction model. The system designs information transmission interfaces between modules to ensure data format conversion and semantic alignment between different computational paradigms. An iterative optimization framework is established, using multiple rounds of generation, verification, and correction loops to gradually improve prediction quality. Each iteration adjusts network parameters, updates constraint weights, or optimizes symbolic inference rules based on the results of the previous round. An adaptive termination mechanism is implemented, judging whether a satisfactory solution quality has been achieved based on indicators such as constraint satisfaction, prediction accuracy improvement, and computational resource consumption, avoiding unnecessary over-optimization. The system also establishes a model performance monitoring mechanism to track the execution efficiency and result quality of each component in real time, promptly identifying and handling potential performance degradation or anomalies.

[0187] In the temporal logic constraint satisfaction verification phase, the system establishes a comprehensive constraint satisfaction verification framework to ensure that the final hybrid prediction model strictly satisfies all predefined temporal logic constraints. The system employs formal verification methods, using mathematical proofs to verify the correctness and completeness of the model output, ensuring no constraint violations are missed. The system establishes a constraint satisfaction statistical analysis mechanism, calculating statistical indicators such as the satisfaction rate, violation frequency, and violation severity of various constraints, providing a quantitative basis for model performance evaluation. The system implements boundary case testing functionality, generating test cases under various extreme and boundary conditions to verify the model's constraint satisfaction capability under difficult conditions. The system also establishes an interpretability analysis for constraint satisfaction, clearly explaining why a certain prediction result satisfies or violates a specific constraint, providing guidance for model improvement and optimization.

[0188] During the model deployment and real-time application phase, the system deploys a fully validated hybrid prediction model to the actual power system operating environment, achieving real-time power prediction that meets time logic constraints. The system features a highly efficient model inference engine, optimizing the computational process and data structure to ensure stable model operation under stringent real-time requirements. A model update mechanism is established to periodically update neural network parameters, symbolic inference rules, and constraint weights based on new operating data and changes in constraints, maintaining the model's adaptability and accuracy. A fault recovery mechanism is implemented, automatically switching to standby or degraded service mode when a component of the model fails, ensuring continuous system availability. Furthermore, a performance evaluation and feedback mechanism is established to continuously collect performance data from practical applications, providing valuable practical experience and a data foundation for further model improvement and optimization.

[0189] In the application of neural symbolic acceleration technology, the system innovatively combines mixed-integer linear programming (MILP) motion planning with neural networks to construct a highly efficient neural symbolic acceleration framework. This framework models the power prediction problem as a constrained optimization problem, with constraints including the aforementioned physical limitations and temporal logic constraints. The system first generates preliminary prediction results through a neural network as the initial solution to the optimization problem; then, it optimizes the initial solution using an efficient MILP solver to ensure that the final result satisfies all constraints. To improve computational efficiency, the system adopts an incremental solution strategy, first solving the global optimization problem at a coarse-grained time scale, and then making local adjustments at a fine-grained scale. The system also utilizes a symbolic inference engine to verify and correct the optimization results, checking for any predictions that violate physical laws and making adjustments according to predefined rules. Through this neural symbolic fusion method, the system maintains the powerful expressive and generalization capabilities of neural networks while ensuring the physical feasibility of the prediction results, significantly improving prediction accuracy and reliability.

[0190] Step 3: Optimize the scheduling strategy for the Blue Bar, including:

[0191] Step 3.1: Based on the renewable energy power generation prediction results, establish a multi-objective dispatch model that comprehensively considers economic efficiency and security, and use a weighted sum method to unify minimizing power purchase cost, maximizing renewable energy absorption rate, and minimizing system frequency deviation into a single optimization objective;

[0192] Specifically, a multi-objective scheduling model considering both economic efficiency and security is established, using a weighted sum method to unify multiple objective functions into a single optimization objective. The objective function includes three core indicators: minimizing electricity purchase cost, maximizing renewable energy absorption rate, and minimizing system frequency deviation, with weighting coefficients set to 0.5, 0.3, and 0.2, respectively. Electricity purchase cost is calculated using time-of-use pricing, divided into three time periods based on grid load characteristics: peak hour price of 0.8 yuan / kWh (8:00-11:00, 18:00-23:00), normal hour price of 0.5 yuan / kWh (7:00-8:00, 11:00-18:00), and off-peak price of 0.3 yuan / kWh (23:00-7:00). The renewable energy absorption rate is defined as the ratio of actual absorption to generateable capacity, with a target value set above 95%. When the absorption rate falls below 90%, a wind and solar curtailment penalty mechanism is triggered, with a penalty cost of 0.2 yuan / kWh.

[0193] In the multi-objective scheduling model construction phase, the system establishes a comprehensive multi-objective scheduling model that considers both economic efficiency and security, based on the renewable energy power generation prediction results. This model comprehensively considers three core optimization objectives: minimizing electricity purchase costs, maximizing renewable energy absorption rate, and minimizing system frequency deviation. To unify the processing of these objectives with different dimensions and priorities, the system employs a weighted sum method for objective fusion, assigning a weight of 0.5 to electricity purchase costs, 0.3 to renewable energy absorption rate, and 0.2 to system frequency deviation. This weighting configuration reflects the relative importance of economic benefits, environmental benefits, and system security. Regarding electricity purchase cost calculation, the system adopts a time-of-use pricing mechanism, dividing the day into three time periods based on grid load characteristics: peak hours (8:00-11:00, 18:00-23:00) with a price of 0.8 yuan / kWh, normal hours (7:00-8:00, 11:00-18:00) with a price of 0.5 yuan / kWh, and off-peak hours (23:00-7:00) with a price of 0.3 yuan / kWh. This time-of-use pricing design effectively guides the shift of electricity load from peak to off-peak hours, smoothing the system load curve. Regarding renewable energy consumption, the system defines the consumption rate as the ratio of actual consumption to generating capacity, setting a target value of over 95%. A penalty mechanism for wind and solar curtailment is introduced; when the consumption rate falls below 90%, a penalty cost of 0.2 yuan will be incurred for each kilowatt-hour of unconsumed renewable energy. This mechanism effectively incentivizes the system to maximize the utilization of clean energy.

[0194] Step 3.2: Construct a complete system constraint framework, including power balance constraints, equipment operation constraints, network security constraints, and unit ramp-up constraints, to ensure the feasibility and safety of the scheduling scheme;

[0195] Specifically, a complete system constraint framework is constructed to ensure the feasibility and safety of the dispatching scheme. Power balance constraints require that the total power generation at any given time equals the sum of load demand, network losses, and energy storage charging, with a balance error controlled within ±2%. Equipment operation constraints include limiting the energy storage system's SOC range to 20%-80%, charging and discharging power not exceeding 90% of rated power, and continuous charging and discharging time not exceeding 4 hours. Network security constraints require that line power flow not exceed 85% of the thermal stability limit, bus voltage maintained within ±7% of the rated value, and system reserve capacity not less than 10% of the maximum load. Furthermore, unit ramping constraints must be met, with the ramping rate of conventional units limited to 5% of rated power per minute, and the energy storage system response time not exceeding 100 milliseconds.

[0196] In the system constraint construction phase, a complete constraint system was established to ensure that the dispatching scheme is physically feasible and safe and reliable. First, there is the power balance constraint, requiring that the total power generation (including renewable energy generation, conventional unit generation, and external power purchase) at any given time must equal the sum of load demand, network losses, and energy storage charging capacity, with an allowable balance error controlled within ±2%. This constraint is a fundamental requirement for the stable operation of the power system. Second, there are equipment operation constraints, including limiting the state of charge (SOC) of the energy storage system to the range of 20%-80% to extend battery life, ensuring that the charging and discharging power does not exceed 90% of the rated power to reserve a safety margin, and limiting continuous charging and discharging time to no more than 4 hours to prevent overuse. Third, there are network security constraints, requiring that the line power flow does not exceed 85% of the thermal stability limit to prevent line overload, that the bus voltage is maintained within ±7% of the rated value to ensure power quality, and that the system reserve capacity is not less than 10% of the maximum load to cope with emergencies. Fourth, there are unit ramp-up constraints, limiting the output change rate of conventional units to no more than 5% / minute of rated power to protect equipment, and ensuring that the energy storage system response time does not exceed 100 milliseconds to meet rapid adjustment requirements. These constraints together constitute a comprehensive constraint system, ensuring that the dispatching scheme meets the operational limitations of the physical system while guaranteeing the safe and stable operation of the power grid.

[0197] Step 3.3: The uncertainty of renewable energy output is handled by using the linear one-dimensional projection constraint sub-Bruker control method. The uncertainty set is constructed by using the Wasserstein distance metric to measure the difference between the actual distribution and the empirical distribution, and the optimal scheduling scheme is obtained.

[0198] In the application of the distributed bar control method, the system innovatively employs a linear one-dimensional projection constraint-based distributed bar control method to handle the uncertainty of renewable energy output. Unlike traditional deterministic optimization methods, this method does not rely on a single point prediction result but considers the probability distribution characteristics of the predicted values, enabling the construction of a reliable uncertainty model based on limited historical data. The system uses the Wasserstein distance to measure the difference between the true and empirical distributions. This distance metric offers better stability than the traditional KL divergence and is particularly suitable for handling extreme cases at the tails of the distribution. By setting an appropriate Wasserstein radius (0.1 in this system), the system achieves a good balance between robustness and conservatism, avoiding both excessive aggressiveness that increases system risk and excessive conservatism that reduces economic efficiency. The distributed bar optimization model constructed based on this method can still meet the system's safe operation requirements in the worst-case scenario while maintaining good economic efficiency, providing a strong guarantee for the safe and economical operation of the power system under conditions of high renewable energy penetration.

[0199] Specifically, step 3.3 includes:

[0200] Step 3.3.1: Based on the renewable energy power generation prediction results, the uncertainty set is defined as a confidence interval centered on the predicted value. The uncertainty interval for wind power is set to ±25% of the predicted value, and the uncertainty interval for photovoltaic power is set to ±20% of the predicted value, thus generating the initial uncertainty set.

[0201] In the uncertainty set definition stage, the system constructs a confidence interval centered on the predicted value based on the renewable energy power generation forecast results, serving as the initial characterization of uncertainty. Considering the differences in uncertainty characteristics between wind power and photovoltaic (PV) power generation, the system sets different uncertainty intervals for each: the uncertainty interval for wind power is set at ±25% of the predicted value, taking into account the greater randomness and prediction difficulty of wind power output; the uncertainty interval for PV is set at ±20% of the predicted value, a relatively narrower interval reflecting the higher predictability of PV power generation under clear weather conditions. These initial interval settings are based on extensive historical data analysis and can cover approximately 90% of actual power generation scenarios. The system also dynamically adjusts the uncertainty intervals for different weather conditions; for example, it expands the uncertainty interval for wind power to ±35% under severe convective weather and expands the uncertainty interval for PV to ±30% under cloudy or overcast weather. Through this adaptive uncertainty modeling based on predicted values ​​and environmental conditions, the system can more accurately characterize the random characteristics of renewable energy output, providing a more reliable uncertainty set for subsequent robust optimization.

[0202] Step 3.3.2: The difference between the true distribution and the empirical distribution of the initial uncertainty set is measured by the Wasserstein distance. The Wasserstein radius is set to 0.1. Based on the statistical characteristics of historical prediction errors, the confidence level is determined to be 95%, and a split-bar uncertainty set is generated.

[0203] In the Wasserstein distance metric stage, the system uses Wasserstein distance to measure the difference between the true distribution of the initial uncertainty set and the empirical distribution based on historical data. The Wasserstein distance essentially measures the minimum "work" required to transform one probability distribution into another, and its advantage lies in providing a meaningful distance metric even if the support sets of the two distributions do not overlap. The system sets the Wasserstein radius to 0.1, a parameter determined through cross-validation, achieving a good balance between model conservatism and economy. Based on the statistical characteristics of historical prediction errors, the system determines a confidence level of 95%, meaning the constructed uncertainty set can cover 95% of possible actual power output scenarios. The system also implements an adaptive Wasserstein radius adjustment mechanism based on the rolling time domain, appropriately reducing the radius to 0.08 during periods of high prediction accuracy (such as midday for photovoltaic power generation) and increasing the radius to 0.15 during periods of greater prediction difficulty (such as the passage of wind fronts). Through this refined uncertainty modeling, the system generates a distributed uncertainty set that reflects the characteristics of the true distribution while possessing statistical guarantees, providing a reliable foundation for subsequent optimization.

[0204] Step 3.3.3: In the set of uncertainties of the sub-Bruker, the sub-Bruker optimization problem is transformed into a deterministic mixed-integer linear programming problem using duality theory, eliminating the complexity of probabilistic constraints and ensuring that the system's safe operation requirements are still met in the worst case, thus obtaining the optimal scheduling scheme.

[0205] Duality theory is one of the core theories in mathematical optimization. It simplifies the solution complexity and provides profound mathematical insights by constructing a dual problem for the primal optimization problem. In the context of robust optimization, duality theory can transform complex optimization problems involving infinite-dimensional probability distributions into finite-dimensional deterministic optimization problems, thereby significantly reducing computational complexity and improving solution efficiency. Through this transformation, the robust optimization problem that originally needed to consider all possible probability distributions becomes a mixed-integer linear programming problem that can be handled by standard optimization solvers, maintaining both the robustness of the solution and ensuring computational feasibility.

[0206] The core idea of ​​duality theory is to transform constrained optimization problems into unconstrained optimization problems using Lagrange duality, by introducing dual variables to handle the constraints in the primal problem. Within the framework of bilabial optimization, the inner layer of the primal problem is a maximization problem involving all possible distributions, while the outer layer is a minimization problem involving decision variables. This two-layer structure makes direct solution extremely difficult. Duality theory transforms the two-layer optimization problem into a single-layer optimization problem by changing the order of maximization and minimization, while using strong duality to ensure that the optimal value of the transformed problem is equal to that of the primal problem. For uncertain sets with Wasserstein distance constraints, the dual problem possesses special structural properties and can be further transformed into a linear or piecewise linear form, laying the foundation for subsequent integer programming modeling.

[0207] The detailed implementation steps are as follows:

[0208] The first stage involves standardizing the sub-Blule bar optimization problem. The system reformulates the original scheduling optimization problem into a standard sub-Blule bar optimization form. The system defines decision variables including the output plans of various power generation devices, the charging and discharging strategies of energy storage systems, and the demand response adjustment schemes. These variables constitute a complete description of the scheduling scheme. The objective function is designed to minimize the expected operating cost under the worst-case scenario, including a weighted combination of generation costs, start-up and shutdown costs, energy storage loss costs, and penalty costs. Constraints include power balance constraints, equipment capacity constraints, network security constraints, and system stability constraints that must be satisfied under all possible renewable energy output distributions. The system pays special attention to handling probabilistic constraints, transforming stochastic constraints such as "the probability of system load failure does not exceed one in a thousand" into sub-Blule bar constraints, ensuring that constraints are satisfied under all distributions within the uncertainty set.

[0209] In the Lagrange duality problem construction phase, the system introduces corresponding Lagrange multipliers for each constraint of the degenerate bar optimization problem, constructing a complete Lagrange function. The system identifies a two-layer structure in the original problem: the inner layer is a maximization problem of all possible distributions within the uncertainty set, and the outer layer is a minimization problem of the decision variables. By applying Lagrange duality theory, the system interchanges the order of optimization between the inner and outer layers, transforming the original "minimize-maximize" problem into a "maximize-minimize" problem. The key to this transformation lies in proving the validity of strong duality. The system ensures that the duality gap is zero by verifying the Slater condition or other sufficient conditions, guaranteeing that the optimal value of the transformed problem is equal to that of the original problem. The system also needs to handle potential unbounded or infeasible cases in the dual problem, ensuring the good properties of the dual problem by adding appropriate regularization terms or constraints.

[0210] In the derivation of the Wasserstein dual representation, the system utilizes the dual representation theory of Wasserstein distance to transform complex constraints involving probability measures into simple constraints in the function space. The dual representation of Wasserstein distance transforms optimization that originally needed to be performed in the probability measure space into optimization in the function space, significantly simplifying the mathematical structure of the problem. Based on Kantorovich duality theory, the system represents the Wasserstein distance as a maximization problem over a class of functions satisfying the Lipschitz conditions; these functions are called dual functions or transport potential functions. By selecting appropriate function approximation methods, such as piecewise linear functions or polynomial functions, the system can discretize the infinite-dimensional function optimization problem into a finite-dimensional linear programming problem. This discretization process requires careful balancing of approximation accuracy and computational complexity; the system employs an adaptive mesh refinement method to optimize the discretization effect.

[0211] In the mixed-integer linear programming modeling phase, the system further organizes the optimization problem, after dual transformation, into a standard mixed-integer linear programming form for solution using mature commercial solvers. The system handles the nonlinear components of the problem, transforming nonlinear constraints into linear constraints through techniques such as piecewise linear approximation, introduction of auxiliary variables, and the Big-M method. For logical constraints involving equipment start-up and shutdown, the system introduces binary variables to represent the operating state of the equipment and establishes corresponding logical constraint relationships. The system pays special attention to the sparsity and special structure of the constraint matrix, improving the stability and efficiency of numerical solutions by rearranging the order of variables and constraints. The system also implements a dynamic problem-scale adjustment mechanism, automatically adjusting the discretization precision and the number of variables based on the actual system scale and computational resource limitations, achieving a balance between solution quality and computation time.

[0212] In the numerical solution and algorithm optimization stages, the system employs a high-performance mixed-integer linear programming solver to handle the transformed deterministic optimization problem. The system selects appropriate solution algorithms, including branch and bound, cutting plane methods, and heuristic algorithms, choosing the most suitable algorithm combination based on the specific structure of the problem. The system implements automatic parameter tuning, automatically setting solver parameters, including convergence tolerance, time limits, memory limits, and parallel computing settings, based on historical solving experience and problem characteristics. To handle large-scale problems, the system uses decomposition algorithms such as Benders decomposition or column generation algorithms to break down large problems into multiple smaller subproblems for solution. The system also implements a solution quality assessment mechanism, evaluating the optimization level of the current solution by calculating the difference between the dual bound and the primal bound, and providing detailed statistical information about the solution process.

[0213] In the worst-case verification and robustness analysis phase, the system performs comprehensive robustness verification on the obtained optimal scheduling scheme to ensure that the scheme maintains feasibility and near-optimal performance under various possible uncertainties. The system constructs a worst-case analysis framework, generating multiple extreme scenarios at the boundaries of the uncertainty set to verify the performance of the scheduling scheme under these scenarios. The system calculates the robustness indicators of the scheme, including key indicators such as performance loss in the worst case, the probability of constraint violation, and system safety margin. When insufficient robustness is found in the scheme, the system automatically adjusts the parameters of the uncertainty set or the weights of the optimization objective, and re-solves until a satisfactory robust scheme is obtained. The system also establishes a sensitivity analysis mechanism to assess the impact of changes in key parameters on the optimal solution, providing operators with guidance on scheme adjustments and risk warnings.

[0214] Specifically, a sub-Bruker optimization model is constructed with the goal of ensuring system safety and economic efficiency even in the worst-case scenario. Duality theory is used to transform the original sub-Bruker optimization problem into a deterministic mixed-integer linear programming problem, eliminating the complexity of probabilistic constraints. The commercial solver Gurobi is used, with a solution accuracy of 0.01%, a maximum number of iterations of 10,000, and a solution time strictly controlled within 30 seconds to meet real-time scheduling requirements. When the solution time exceeds 25 seconds, the solution accuracy is automatically reduced to 0.1% to ensure timely output of the scheduling scheme. An emergency mechanism for solution failure is established: when the optimization problem has no feasible solution, some non-critical constraints are automatically relaxed to prioritize system safety.

[0215] In the sub-Bruker optimization problem-solving phase, the system, based on the constructed sub-Bruker uncertainty set, utilizes duality theory to transform the original sub-Bruker optimization problem into a computationally tractable deterministic problem. The core of this transformation process is to convert the infinite-dimensional optimization problem of "taking the worst-case scenario for all possible distributions" in the original problem into a finite-dimensional deterministic problem through strong duality. The transformed problem is a mixed-integer linear programming (MILP) problem, solved using the commercial solver Gurobi, with a solution accuracy of 0.01% and a maximum of 10,000 iterations. To meet the real-time scheduling requirements, the system strictly controls the solution time to no more than 30 seconds. When the solution time is detected to be close to 25 seconds, the solution accuracy is automatically reduced to 0.1% to ensure timely output of the scheduling scheme. The system also establishes an emergency mechanism for solution failures. When the optimization problem has no feasible solution, an approximate feasible solution is found by gradually relaxing non-critical constraints (such as temporarily expanding the energy storage SOC range to 15%-85%), while prioritizing key constraints related to system safety. Through this bibliometric optimization method, the optimal scheduling scheme generated by the system can maintain feasibility and near-optimality under various possible renewable energy output conditions, significantly improving the power system's adaptability to uncertainty.

[0216] Step 4: Distributed real-time scheduling based on consensus algorithms, including:

[0217] Step 4.1: Based on the optimal scheduling scheme, model the energy storage system, controllable load, and distributed power source in the power system as intelligent agents with autonomous decision-making capabilities. Each agent is equipped with an independent computing unit and communication module.

[0218] Specifically, various devices in the power system (energy storage systems, controllable loads, distributed generation, etc.) are modeled as intelligent agents with autonomous decision-making capabilities. Each agent is equipped with an independent computing unit and communication module. The local status information maintained by the agents includes: the state of charge (SOC), charge and discharge efficiency, cycle count, temperature, and other operating parameters of the energy storage system; the current power level, adjustable range, response time, and user comfort constraints of the controllable load; and the real-time power generation, equipment health status, and weather forecast information of the distributed generation. Control variables cover the charge and discharge power commands of the energy storage system, the adjustment amounts of the controllable load, and the output adjustment commands of the distributed generation. All variables are stored in a standardized format and updated periodically.

[0219] In the intelligent agent modeling phase, the system abstracts various devices in the power system into intelligent agents with autonomous decision-making capabilities based on the optimal scheduling scheme, realizing the basic architecture for distributed coordinated control. The energy storage system agent manages the charging and discharging behavior of batteries, maintaining key operating parameters such as state of charge (SOC), charging and discharging efficiency, cycle count, and battery temperature. The controllable load agent manages demand-side response resources, recording information such as current power levels, adjustable range, response time, and user comfort constraints. The distributed power agent monitors renewable energy generation equipment, tracking real-time power generation, equipment health status, and local weather conditions. Each agent is equipped with an independent computing unit, using an embedded processor or industrial PC to run lightweight decision-making algorithms; it also has a communication module supporting multiple communication protocols such as Modbus, IEC 61850, and DNP3, ensuring compatibility with equipment from different manufacturers. Internally, the agents adopt a hierarchical control architecture: the bottom layer executes basic device control logic, the middle layer implements local optimization decisions, and the top layer is responsible for coordinating with other agents. This intelligent agent architecture breaks down system complexity into manageable modules, with each agent focusing on optimizing and controlling its own device while achieving global coordination through information exchange, greatly improving the system's scalability and robustness.

[0220] Step 4.2: Design a distributed protocol based on the push-pull average consensus algorithm, in which each agent periodically exchanges information with a predefined set of neighbor agents, wherein the neighbor relationship is determined based on the physical connection topology and communication network structure;

[0221] Specifically, a distributed protocol based on the Push-Pull Average Consensus with Dynamic Compression (PP-ACDC) algorithm is designed, which has the advantages of fast convergence and low communication complexity. Each agent periodically exchanges information with a predefined set of neighboring agents, and the neighbor relationships are determined based on the physical connection topology and communication network structure.

[0222] In the distributed protocol design phase, the system employs a push-pull average consensus algorithm to achieve efficient coordination among multiple agents. The core idea of ​​this protocol is to achieve global consensus through local information exchange, eliminating the need for a central controller and resulting in extremely high system reliability and scalability. The system first defines neighbor relationships based on physical network topology and communication reachability. Each agent exchanges information only with its direct neighbors, significantly reducing communication burden. Neighbor relationships are represented using a weighted undirected graph, where edge weights reflect the physical connection strength and communication quality between nodes. An adaptive weight adjustment mechanism allows the system to dynamically respond to changes in communication link quality. The protocol design specifically considers communication latency and packet loss, employing timeout retransmission and message acknowledgment mechanisms to ensure the reliability of information exchange. To improve the convergence speed of the protocol, the system adopts accelerated consensus technology, significantly reducing the number of iterations required to reach consensus by optimizing the iteration step size and utilizing historical information to predict future states. The protocol also includes an anomaly detection mechanism. When an agent's behavior is detected as abnormal or inconsistent information is provided, its influence weight in the consensus process is automatically adjusted to prevent malicious or faulty nodes from disrupting the overall consensus. This distributed protocol design enables the system to achieve globally optimal control through local communication without a central controller, greatly improving the system's reliability and robustness.

[0223] Step 4.3: During the push operation phase, each agent sends its current state vector and gradient information to all neighboring agents, and the sending weights are dynamically adjusted according to the degree centrality of the network topology;

[0224] Specifically, in the push phase, agent i sends its current state vector xi and gradient information gi to all neighboring agents, and the sending weights are dynamically adjusted according to the degree centrality of the network topology.

[0225] In the push operation phase, the system employs an efficient information push mechanism for each agent to proactively disseminate critical state information. At the start of each communication cycle, the agent first collects local sensor data and updates its internal state, including device operating parameters, environmental conditions, and control objectives. Then, the agent calculates the difference between the current state vector and the state of the previous cycle, generating incremental update information. This incremental transmission strategy significantly reduces the amount of communication data. Simultaneously, the agent calculates the gradient information of state changes, reflecting parameter change trends and helping the receiver predict future state changes. After determining the content to be sent, the agent dynamically adjusts the transmission weight based on the degree centrality of the network topology. Core nodes (nodes with high connectivity) receive higher weights, while edge nodes receive lower weights. This differentiated weight allocation accelerates the diffusion of information throughout the network. The system also implements a priority transmission mechanism, giving higher priority to critical state changes (such as energy storage SOC approaching its limit or sudden changes in renewable energy output), ensuring timely delivery of important information. Through this carefully designed push operation mechanism, the system achieves efficient dissemination of critical information, laying the foundation for subsequent consensus formation.

[0226] Step 4.4: During the pull operation phase, each agent receives state information from its neighbors and updates its local state by merging the neighbor information using a weighted average method;

[0227] Specifically, during the pull operation phase, agent i receives state information from its neighbors and updates its local state by merging the neighbor information using a weighted average: x i (k+1) = x i (k) + α∑w ij [x j (k) - x i [(k)], where α is the learning rate, w ij This represents the neighbor weight coefficient.

[0228] In the pull operation phase, the system employs an intelligent information fusion mechanism, enabling each agent to effectively integrate state information from its neighbors. Upon receiving state vectors and gradient information from neighbors, the agent first verifies data validity, checking timestamps to ensure timeliness, verifying digital signatures to ensure data source credibility, and checking numerical ranges to ensure data rationality. Verified data then enters the fusion processing phase. The system uses a weighted average method to integrate multi-source information, with weight coefficients dynamically adjusted based on data credibility, neighbor node importance, and historical information accuracy. To handle potential conflicts between different neighbor information, the system implements a conflict resolution mechanism based on evidence theory, enabling reasonable decisions even with contradictory information. After information fusion, the agent updates its local state, calculates the deviation between the new state and the target state, and adjusts its control strategy accordingly. The system also implements a learning-reinforcement fusion mechanism, continuously optimizing fusion weights through reinforcement learning, allowing agents to gradually identify and rely more heavily on neighbors providing high-quality information. This intelligent pull operation mechanism allows each agent to fully utilize distributed information within the network, achieving global coordination while maintaining autonomous decision-making capabilities.

[0229] Step 4.5: Adopt an adaptive quantization scheme;

[0230] Specifically, an information-theory-based dynamic compression mechanism is introduced. Information importance is assessed based on Shannon entropy, and quantization precision is adaptively adjusted to reduce communication overhead. A three-level quantization system is established: key control signals (such as energy storage charging / discharging commands and emergency load shedding commands) use 16-bit high-precision quantization, with quantization error controlled within 0.01%; status monitoring information (such as voltage, current, and power measurements) uses 8-bit medium-precision quantization, with quantization error controlled within 0.1%; and auxiliary information (such as environmental parameters like temperature and humidity) uses 4-bit low-precision quantization, with quantization error allowed within 1%. An adaptive quantization scheme is designed to monitor the variance and rate of change of the system's operating status in real time. When the system is in steady state (variance < 0.05), a coarse quantization strategy is used to reduce communication overhead by 50%; when system disturbances are detected (rate of change > 5% / minute), it automatically switches to fine quantization mode to ensure control accuracy.

[0231] In the implementation of the adaptive quantization scheme, the system developed a dynamic compression mechanism based on information theory. This mechanism adaptively adjusts the quantization precision according to the importance of the information, significantly reducing the communication burden. The mechanism first assesses the importance of information by calculating its Shannon entropy. A high entropy value indicates that the information contains more uncertainty and requires more refined quantization; a low entropy value indicates that the information changes less and can be coarsely quantized. Based on this principle, the system established a three-level quantization system: key control signals such as energy storage charging and discharging commands and emergency load shedding commands use 16-bit high-precision quantization, with the quantization error strictly controlled within 0.01%, ensuring control accuracy; status monitoring information such as voltage, current, and power measurements use 8-bit medium-precision quantization, with the quantization error controlled within 0.1%, balancing accuracy and bandwidth requirements; auxiliary information such as environmental parameters such as temperature and humidity uses 4-bit low-precision quantization, with the quantization error allowed within 1%, significantly saving communication resources. The system also implements a context-aware adaptive quantization strategy, monitoring the variance and rate of change of the operating status in real time. When the system is in steady state (variance less than 0.05), it automatically switches to coarse quantization mode, reducing communication overhead by up to 50%. When a system disturbance is detected (rate of change exceeding 5% / minute), it immediately switches to fine quantization mode to ensure control accuracy. This intelligent quantization scheme significantly reduces the communication burden while ensuring control performance, and improves the system's scalability and response speed.

[0232] Specifically, step 4.5 includes:

[0233] Step 4.5.1: The information-theory-based dynamic compression mechanism assesses the importance of information based on Shannon entropy and adaptively adjusts the quantization precision to reduce the communication burden;

[0234] In the information importance assessment stage, the system constructs a refined information value assessment mechanism based on information theory principles, providing a decision-making basis for adaptive quantification. First, the system quantifies the uncertainty and importance of information by calculating its Shannon entropy. For each state variable x to be transmitted, its probability distribution p(x) is calculated, and the entropy value is obtained accordingly. A high entropy value indicates that the information contains more uncertainty and requires more refined quantification; a low entropy value indicates that the information changes less or is highly predictable, and coarse quantification can be used. The system also introduces the concept of conditional entropy, considering the correlation between variables to avoid transmitting redundant information. In addition to the basic entropy calculation, the system also considers three dimensions for comprehensive evaluation: timeliness, relevance, and control sensitivity. Timeliness assessment is based on the difference between the data's generation time and the current time, using an exponential decay function to quantify timeliness, ensuring that the latest information receives higher priority. Relevance assessment analyzes the correlation between variables and key control objectives, using Pearson correlation coefficient and mutual information as dual indicators; highly correlated variables receive higher importance scores. Control sensitivity assessment analyzes the impact of variable changes on the system's control performance, determining the control weight of each variable through sensitivity analysis. The system integrates the scores from these three dimensions into a final importance index through a weighted summation method, dynamically adjusts the weights of each dimension to adapt to different operating scenarios, and provides a scientific basis for subsequent accuracy allocation.

[0235] Step 4.5.2: Establish a three-level quantization system. Key control signals use 16-bit high-precision quantization, status monitoring information uses 8-bit medium-precision quantization, and auxiliary information uses 4-bit low-precision quantization.

[0236] In the precision allocation stage, based on the aforementioned information importance assessment results, the system implements an optimal precision allocation strategy under resource-constrained conditions. The system first establishes a precision-bandwidth model, quantifying the data volume and transmission bandwidth requirements corresponding to different precision levels, and considering the bandwidth limitations of different communication channels. Based on this, the system models the precision allocation problem as a constrained optimization problem, aiming to maximize information fidelity, with the constraint that the total bandwidth does not exceed the available bandwidth. The system uses the Lagrange multiplier method to solve this optimization problem, obtaining the optimal number of quantization bits for each variable. To achieve flexible precision control, the system designs a three-level quantization system: key control signals such as power adjustment commands and protection trigger signals use 16-bit high-precision quantization, with quantization errors strictly controlled within 0.01%; status monitoring information such as voltage, current, and power measurements use 8-bit medium-precision quantization, with quantization errors controlled within 0.1%; and auxiliary information such as environmental parameters and non-critical status quantities use 4-bit low-precision quantization, with quantization errors allowed within 1%. The system also implements a dynamic precision adjustment mechanism based on network congestion status. When network congestion is detected, the quantization precision of non-critical information is automatically reduced to ensure the timely transmission of critical information. When the network condition improves, the normal quantization precision is gradually restored to achieve efficient utilization of communication resources.

[0237] Step 4.5.3: Monitor the variance and rate of change of the system's operating state in real time. When the system's operating state is in a steady state, adopt a coarse quantization strategy to reduce communication overhead. When a disturbance is detected in the system's operating state, automatically switch to fine quantization mode to ensure control accuracy.

[0238] In the context-aware dynamic adjustment phase, the system implements an intelligent mechanism capable of sensing changes in the operating environment and dynamically adjusting the quantization strategy. The system continuously monitors key operational indicators, including system state change rate, control error, network latency, and packet loss rate, constructing a real-time profile of the system's operating status. Based on this monitoring data, the system categorizes operating states into four types: stable state, slightly fluctuating state, drastically changing state, and emergency state, and configures corresponding quantization strategies for each state. In a stable state, the system employs a low-frequency, low-precision communication mode, transmitting only information where state changes exceed a threshold, reducing communication volume by up to 70%. In a slightly fluctuating state, the system increases the sampling frequency and quantization accuracy of key variables, ensuring the control system can respond promptly to small disturbances. In a drastically changing state, the system activates a high-frequency, high-precision full-quantization communication mode, ensuring the control system receives complete and accurate state information. In an emergency state, the system prioritizes the transmission of critical control signals, even sacrificing the transmission of non-critical information to ensure the timely delivery of critical information. The system also implements a predictive adjustment mechanism, analyzing historical data and current trends to predict future state changes and adjust the quantization strategy in advance, further enhancing the system's adaptability to environmental changes. This context-aware dynamic adjustment mechanism enables the system to achieve efficient utilization of communication resources while ensuring control performance, significantly improving the scalability and real-time performance of the distributed control system.

[0239] Step 4.6: Establish a multi-level convergence criterion, and achieve scheduling consensus through state convergence, objective function convergence and consistency error criterion, and execute the real-time power adjustment operation.

[0240] Specifically, a multi-level convergence criterion is established to ensure the reliability of the algorithm. The first-level criterion is based on state convergence: when the state change of each agent over 10 consecutive scheduling cycles is ||x||. i (k+1)-x iThe convergence condition is met when (k)‖<0.1%. The secondary criterion is based on the convergence of the objective function: convergence is confirmed when the relative change in the global objective function value is less than 0.05%. The tertiary criterion is based on the consistency error: consensus is reached when the maximum deviation of all agent states is less than a set threshold. Each scheduling cycle is set to 5 minutes, subdivided into 60 communication time slots, each slot being 50 seconds, allowing sufficient time for algorithm convergence. Communication latency is strictly controlled within 100ms, and a time synchronization protocol is used to ensure clock consistency among agents, with clock deviation controlled within ±10ms. A communication fault detection and recovery mechanism is established; when a communication interruption is detected, the agent automatically switches to local autonomous mode, making decisions based on historical data and preset rules to ensure continuous system operation.

[0241] In the convergence criterion establishment phase, the system constructs a multi-level convergence evaluation system to ensure that the distributed algorithm can reliably achieve global consensus. The first-level criterion is based on state convergence, monitoring the absolute and relative magnitudes of state changes for each agent. When the state changes of all agents are less than a preset threshold (typically 0.1%) for 10 consecutive scheduling cycles, the state is considered stable. The second-level criterion is based on objective function convergence, calculating the changing trend of the global optimization objective function value. When the relative change is less than 0.05% for 5 consecutive cycles, the optimization process is considered convergent. The third-level criterion is based on consistency error, calculating the maximum and average deviations of all agent states. When the maximum deviation is less than a set threshold and remains stable, consensus is confirmed. The system assigns different weights to these three levels of criteria. When the comprehensive score exceeds the threshold, a consensus confirmation mechanism is triggered, and each agent locks its current state and executes corresponding control actions. To ensure the algorithm converges within a finite time, the system sets a maximum iteration limit. When the number of iterations exceeds the limit but convergence is not yet complete, an approximate consensus strategy is adopted, prioritizing the consistency of key control objectives. The system also implements a hierarchical convergence mechanism, allowing different control levels to converge at different speeds. Emergency control actions can reach consensus and be executed in a shorter time, while non-critical adjustments can have a longer negotiation period. This multi-level convergence guarantee mechanism enables the distributed control system to reliably reach consensus and execute coordinated control actions under various network conditions.

[0242] Step 5: Closed-loop feedback and adaptive optimization, including:

[0243] Step 5.1: Establish a real-time data feedback channel and collect the actual operating status of the equipment through the SCADA system and smart meters, including the actual charging and discharging power of the energy storage system, the actual adjustment of the controllable load, and the actual output of renewable energy.

[0244] Specifically, a real-time data feedback channel is established to collect the actual execution status of the equipment through the SCADA system and smart meters, including the actual charging and discharging power of the energy storage system, the actual adjustment of the controllable load, and the actual output of renewable energy, with a data update frequency of 1 minute.

[0245] In the real-time data feedback channel establishment phase, the system constructs a comprehensive data acquisition and feedback mechanism to provide real-time data support for closed-loop control. First, the system collects operating status data of power equipment through the SCADA system, including parameters such as active power, reactive power, speed, and temperature of generator sets, with a sampling frequency on the order of seconds. It also collects electricity consumption data from the user side through smart meters, including power, voltage, current, and power factor, with a sampling frequency on the order of minutes. Furthermore, it collects environmental parameters such as wind speed, light intensity, temperature, and humidity through a dedicated sensor network, with a sampling frequency on the order of minutes. All collected data is accompanied by precise timestamps, and the entire system clock is synchronized via the NTP protocol, with a time synchronization accuracy better than 1 millisecond, ensuring data time sequence consistency. Data transmission employs a multi-level caching mechanism. Local devices first store data in the cache of edge computing nodes, and then periodically upload it in batches to the central database, ensuring both real-time data transmission and improved transmission efficiency. The system also implements a data quality monitoring mechanism, using triple verification of validity, consistency, and integrity checks to identify and mark potential data anomalies, ensuring the reliability of the feedback data. To handle communication interruptions, the system incorporates a data recovery mechanism. Once communication is restored, historical data from the period of interruption is automatically retransmitted, ensuring data integrity. This comprehensive and reliable real-time data feedback channel provides a solid data foundation for closed-loop system optimization.

[0246] Step 5.2: Calculate the prediction error index, which includes mean absolute error, root mean square error and mean absolute percentage error. When any index exceeds a preset threshold, the model retraining mechanism is triggered.

[0247] Specifically, prediction error metrics are calculated, including mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE). When any metric exceeds a preset threshold (MAE>5%, RMSE>8%, MAPE>10%), the model retraining mechanism is triggered.

[0248] In the prediction error analysis and model update stages, the system continuously evaluates the performance of the prediction model and performs adaptive optimization based on real-time feedback data. The system calculates various prediction error metrics, including Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE), reflecting the absolute deviation, volatility, and relative error of the prediction, respectively. The system calculates these metrics by time scale (hourly, daily, weekly) and prediction object (wind power, solar power, load), forming a multi-dimensional performance evaluation matrix. When any metric exceeds a preset threshold (e.g., MAPE exceeds 15%), a model retraining mechanism is triggered. To improve model update efficiency, the system adopts an incremental learning approach, eliminating the need to retrain the entire model; instead, it adjusts model parameters based on new samples. The weight of new samples is set to 0.1, and the weight of historical samples is adjusted according to an exponential decay law with a decay factor of 0.95. This weight setting allows the model to gradually adapt to new data patterns while maintaining memory of historical patterns. The system also implements a model structure adaptive mechanism, automatically adjusting hyperparameters such as the number of network layers, neurons, and learning rate through Bayesian optimization to find the most suitable model structure for the current data characteristics. To prevent overfitting, the system employs an early stopping strategy and regularization techniques to ensure the model's generalization ability. Through this continuous prediction error analysis and model update mechanism, the system can continuously improve prediction accuracy and adapt to dynamic changes in power system and environmental conditions.

[0249] Step 5.3: Design an online learning algorithm to update the prediction model parameters using incremental learning. The updated sample weight is set to 0.1, and the historical sample weight is adjusted according to the exponential decay law.

[0250] Specifically, an online learning algorithm is designed, and the prediction model parameters are updated using an incremental learning method. The weight of new samples is set to 0.1, and the weight of historical samples is adjusted according to the exponential decay law, with a decay factor of 0.95.

[0251] In the dispatch performance evaluation phase, the system constructs a comprehensive performance evaluation framework, evaluating the effectiveness of dispatch strategies from three dimensions: economy, safety, and environmental protection. The economy evaluation primarily examines operating costs, including the sum of electricity purchase costs, start-up and shutdown costs, maintenance costs, and penalty costs (such as wind and solar curtailment penalties), calculating cost savings rates by comparing with benchmark strategies. The safety evaluation focuses on system stability indicators, including frequency stability (root mean square value of frequency deviation), voltage stability (maximum voltage deviation at critical nodes), reserve capacity adequacy ratio (ratio of actual reserve to minimum requirement), and N-1 safety verification pass rate. The environmental evaluation focuses on the effectiveness of renewable energy utilization, calculating renewable energy absorption rate (ratio of actual absorption to generateable capacity), carbon emission intensity (carbon emissions per unit of electricity), and energy conservation and emission reduction contribution (emission reduction compared to traditional dispatch methods). The system integrates these indicators into an overall performance score using a weighted approach. When a performance indicator deviates from the target value by more than 5%, an adaptive adjustment mechanism for the dispatch strategy is triggered. The system also implements spatiotemporal analysis of performance indicators, identifying the temporal patterns and spatial distribution of performance fluctuations, providing targeted guidance for strategy optimization. Through this multi-dimensional scheduling performance evaluation, the system can fully grasp the advantages and disadvantages of the scheduling strategy, providing a clear direction for subsequent optimization.

[0252] Step 5.4: Construct a scheduling performance evaluation system, including economic indicators, safety indicators, and environmental indicators. When the performance indicators deviate from the target value by more than 5%, trigger the adaptive adjustment mechanism of the scheduling strategy.

[0253] Specifically, a scheduling performance evaluation system is constructed, including economic indicators (electricity purchase cost, wind and solar curtailment rate), safety indicators (frequency deviation, voltage deviation), and environmental indicators (carbon emissions, renewable energy utilization rate). When the performance indicators deviate from the target value by more than 5%, an adaptive adjustment mechanism for the scheduling strategy is triggered.

[0254] In the reinforcement learning optimization phase, the system utilizes the Deep Deterministic Policy Gradient (DDPG) algorithm to construct an adaptive scheduling strategy optimization framework. This framework models the power system as a Markov decision process. The state space includes key information such as system load, renewable energy output, energy storage status, and grid topology; the action space includes control variables such as conventional unit scheduling commands, energy storage charging and discharging commands, and controllable load adjustment commands; the reward function is designed as a weighted combination of economic, safety, and environmental indicators, with weights dynamically adjusted according to the system's operational objectives. The DDPG algorithm comprises two deep neural networks: the Actor network is responsible for generating the optimal control action based on the current state, employing a four-layer fully connected structure with [256, 128, 64] hidden layer neurons and ReLU activation function; the Critic network is responsible for evaluating the value of state-action pairs, guiding the optimization direction of the Actor network, also employing a four-layer structure, but integrating action information in the second layer. The system uses an experience replay technique, maintaining a replay buffer with a capacity of 10,000 to store historical state-action-reward-next state sequences, and randomly sampling batches of 64 samples for each training iteration. To balance exploration and exploitation, the system employs an Ornstein-Uhlenbeck noise process to generate exploration noise, with noise parameters adaptively adjusted as the learning process progresses. The system also implements a target network mechanism, using slowly updated target Actor and Critic networks to calculate target values, improving training stability. Through this reinforcement learning-based continuous optimization mechanism, the scheduling strategy can continuously adapt to system changes and environmental disturbances, achieving sustained improvement in scheduling performance.

[0255] Step 5.5: Based on the aforementioned scheduling performance evaluation system, the scheduling strategy is continuously optimized using a reinforcement learning algorithm. The system operating state is taken as input, the scheduling action is taken as output, and a multi-objective weighted sum is taken as the reward function to form the adaptive closed-loop control system.

[0256] Specifically, a reinforcement learning algorithm (DDPG) is used to continuously optimize the scheduling strategy, taking the system state as input and scheduling actions as output, with a multi-objective weighted sum as the reward function. The state space includes renewable energy prediction bias, energy storage SOC, load change rate, and electricity price information, while the action space includes energy storage charging and discharging commands, load adjustment commands, and grid interaction power. An experience replay buffer is set up to store historical state-action-reward sequences, with a buffer size of 10,000. Batch learning is performed every 100 steps, with a batch size of 32.

[0257] In the construction phase of the adaptive closed-loop control system, the system integrates all the aforementioned components into a complete adaptive closed-loop control system, realizing full-process closed-loop control from data acquisition, state estimation, decision optimization to execution feedback. The system adopts a hierarchical control architecture. The bottom layer implements basic equipment control and protection functions with a response time in milliseconds; the middle layer implements rule-based rapid adjustment with a response time in seconds; and the top layer implements strategy adjustment based on optimization and learning with a response time in minutes. Controllers at each layer collaborate seamlessly through clear interfaces and protocols. The higher-level controller sets goals and constraints, while the lower-level controllers execute specific control actions. The system achieves coordinated optimization across multiple time scales: short-term scheduling (5-minute level) focuses on system balance and safety constraints; medium-term scheduling (hourly level) optimizes economic efficiency and renewable energy consumption; and long-term scheduling (daily level) considers equipment lifespan and maintenance plans. To cope with emergencies, the system is designed with an event-driven control mode switching mechanism. When a major disturbance or emergency is detected, it automatically switches to a more conservative control strategy, prioritizing system safety. The system also implements control performance monitoring and self-diagnostic functions, continuously evaluating control effectiveness, identifying potential problems, and adjusting control parameters or strategies through a self-healing mechanism. This fully integrated adaptive closed-loop control system can achieve safe, economical, and environmentally friendly coordinated operation in complex and ever-changing power system environments, providing strong technical support for the integration of high proportions of renewable energy.

[0258] Specifically, step 5.5 includes:

[0259] Step 5.5.1: Based on the aforementioned scheduling performance evaluation system, the scheduling strategy is continuously optimized using the deep deterministic policy gradient algorithm, taking the system running state as input, the scheduling action as output, and the multi-objective weighted sum as the reward function;

[0260] In the state space construction phase, the system designs a comprehensive state representation scheme based on the physical characteristics and operational requirements of the power system, providing complete environmental awareness capabilities for reinforcement learning. The state space contains four main categories of information: power system operating status, renewable energy status, energy storage system status, and environmental conditions. Power system operating status includes load curves (current values ​​and 24-hour forecasts), system frequency, critical node voltages, line power flow, and reserve capacity, reflecting the system's basic operating condition and safety margin. Renewable energy status includes actual and predicted output of wind and solar power, along with their uncertainty range. Output forecasts cover the next 24 hours with a 15-minute time resolution, and uncertainty is expressed as the standard deviation of the predicted values. Energy storage system status includes current state of charge (SOC), charge / discharge efficiency (considering temperature effects), cycle count, and health status index. These parameters comprehensively reflect the availability and constraints of the energy storage system. Environmental conditions include meteorological data (temperature, humidity, wind speed, and solar irradiance) and their predicted values, as well as electricity price signals (time-of-use pricing and peak-shaving ancillary service prices). The system standardizes all state variables, mapping them to the [-1,1] interval to eliminate the impact of dimensional differences on learning. To handle the high-dimensional state space, the system employs principal component analysis (PCA) and autoencoder techniques for dimensionality reduction, retaining 95% of the information while reducing computational complexity, enabling the reinforcement learning algorithm to efficiently handle complex system states.

[0261] Step 5.5.2: Construct the state space, including renewable energy forecast bias, energy storage state of charge, load change rate, and electricity price information; the action space includes energy storage charging and discharging commands, load adjustment commands, and grid interaction power.

[0262] In the action space design phase, the system constructs a structured representation of control actions to ensure that the reinforcement learning algorithm can generate physically feasible and effective control strategies. The action space mainly includes three types of control variables: generation-side regulation, energy storage-side regulation, and demand-side response. Generation-side regulation includes output adjustment commands for conventional units, considering the unit's ramp-up capability, minimum technical output, and maximum output limits, with a control granularity of 1% of the unit's rated capacity. Energy storage-side regulation includes charging and discharging power commands and operating mode selection (such as frequency regulation mode, peak shaving and valley filling mode, and economic operation mode), with charging and discharging power limited to within 90% of rated power to extend equipment lifespan. Demand-side response includes interruptible load reduction commands and adjustable load peak-shifting commands, considering user comfort constraints and response time limitations. The system adopts a continuous action space design, using the tanh function to map the neural network output to the [-1,1] interval, and then mapping it to the actual control range through a linear transformation. To ensure the physical feasibility of actions, the system incorporates an action projection layer. This layer projects unsatisfactory original actions onto the nearest feasible solution. This design significantly accelerates the learning process and avoids ineffective exploration of infeasible regions. The system also implements an action smoothing mechanism to limit the variation in control commands between adjacent time periods, preventing frequent and large-scale adjustments from adversely affecting the equipment and improving the practicality and acceptability of the control strategy.

[0263] Step 5.5.3: Set up an experience replay buffer to store the historical state-action-reward sequence. The buffer size is 10,000. Perform batch learning every 100 steps with a batch size of 32.

[0264] In the reward function design phase, the system constructs a multi-objective weighted comprehensive reward mechanism to guide the reinforcement learning algorithm towards the overall optimal direction of the system. The reward function comprises four core components: economic reward, safety reward, environmental reward, and smoothness reward. The economic reward is calculated based on total operating costs, including the weighted sum of electricity purchase costs, start-up and shutdown costs, maintenance costs, and penalty costs (such as wind and solar curtailment penalties), with weights set according to the system's operational objectives. The safety reward considers system stability indicators, including frequency deviation, voltage deviation, and reserve capacity adequacy rate. When these indicators approach the safety boundary, the reward value drops sharply, forming a "cliff-like" penalty, strongly guiding the algorithm to avoid unsafe areas. The environmental reward focuses on the effectiveness of renewable energy utilization, rewarding high absorption rates and low carbon emissions. Specifically, it is calculated as the weighted sum of the positive deviation of the renewable energy absorption rate from the benchmark value and the negative deviation of the carbon emission intensity from the benchmark value. The smoothness reward encourages the continuity and stability of control actions, penalizing frequent and large adjustments, and is calculated as the negative weighted sum of changes in control commands between adjacent time periods. The system adopts an adaptive weighting mechanism, which dynamically adjusts the weight of each part of the reward according to the current system state and operating objectives. For example, it increases the economic weight during peak load periods, increases the environmental weight during periods of high renewable energy generation, and increases the safety weight when the system fluctuates greatly, so that the reinforcement learning algorithm can generate the most suitable control strategy according to the characteristics of the scenario.

[0265] Step 5.5.4: Establish an actor network and a critic network, wherein the actor network outputs continuous action values ​​and the critic network evaluates the value function of state-action pairs;

[0266] In the implementation of the deep deterministic policy gradient algorithm, the system constructs an end-to-end reinforcement learning optimization system based on the DDPG framework. This system comprises four core neural networks: an Actor network, a Critic network, and their corresponding target networks. The Actor network is responsible for policy generation, employing a four-layer fully connected structure. The input layer corresponds to the state dimension, the number of hidden layer neurons is [256, 128, 64], the ReLU activation function is used to introduce non-linearity, and the output layer uses the tanh activation function to map the output to the [-1, 1] interval. The Critic network is responsible for value evaluation, also employing a four-layer structure, but incorporating action information in the second layer, outputting a single Q-value to evaluate the value of the current state-action pair. To improve training stability, the system adopts a target network mechanism. The target network parameters are slowly updated to track the main network parameters via soft updates, with an update rate τ set to 0.001. This design effectively mitigates the instability and overestimation problems of Q-value estimation. The system maintains an experience replay buffer with a capacity of 10000 to store transition samples (s). t , a t , r t , s t+1Each training iteration randomly samples a batch of 64 samples for learning. This experience replay mechanism breaks the temporal correlation between samples, improving learning efficiency and stability. To balance exploration and utilization, Ornstein-Uhlenbeck noise is added to the Actor network output during the training phase. The noise parameters are θ=0.15 and σ=0.2, which gradually decays as the learning process progresses, ensuring that the algorithm can fully explore the action space while gradually converging to the optimal policy.

[0267] Step 5.5.5: Update the actor network parameters to maximize the cumulative reward using the gradient ascent method, and update the critic network parameters using the temporal difference error to form a continuously optimized reinforcement learning strategy.

[0268] In the experience replay and batch learning stages, the system implements efficient sample management and learning mechanisms, improving the data efficiency and training stability of reinforcement learning. The system designs a priority-based experience replay mechanism, calculating sample priority based on the sample's time decay factor and TD error magnitude. Samples with larger TD errors contain more valuable information and receive higher sampling probabilities. To prevent high-priority samples from excessively dominating the training process, the system employs an importance sampling technique, adjusting and updating weights to offset the impact of sampling bias. The system also implements a multi-step reward calculation mechanism, considering not only immediate rewards but also discounted rewards for the next n steps (n=4). This design accelerates the propagation of valuable information and improves learning efficiency. To enhance the algorithm's generalization ability, the system employs an experience generalization technique, generating synthetic samples by adding small random perturbations to the original samples, expanding the training dataset, and improving the model's robustness to noise and uncertainty. The system also implements a distributed experience collection mechanism, with multiple agents interacting with the environment in parallel to collect samples and sharing an experience replay buffer, significantly improving sample collection efficiency. During batch learning, the system employs gradient pruning to prevent gradient explosion and limits the gradient norm to no more than 10; batch normalization accelerates training convergence; and a learning rate decay strategy is used, with an initial learning rate of 0.001 that decreases to 0.9 times every 100,000 training steps, ensuring the algorithm can stably converge to a local optimum. Through these advanced experience replay and batch learning techniques, the system significantly improves the training efficiency and performance of reinforcement learning, enabling high-quality solutions to complex power dispatch optimization problems within a limited training time.

[0269] like Figure 2 As shown, the present invention also provides a renewable energy power generation prediction and power dispatch system, comprising:

[0270] The data acquisition and preprocessing module is used to acquire historical power generation data and real-time meteorological data of renewable energy generation based on a standardized data interface, and to perform linear interpolation to fill missing values ​​and box plot anomaly detection on the historical power generation data and real-time meteorological data to obtain a normalized training dataset.

[0271] The neural symbol fusion prediction module is used to construct a hybrid prediction model based on the normalized training dataset using neural symbol acceleration technology with temporal logic constraints. It captures short-term temporal dependencies through an LSTM network and extracts medium- and long-term spatial-temporal features through a CNN-Transformer network to generate renewable energy power generation prediction results.

[0272] The sub-Blu-bar scheduling optimization module is used to formulate a scheduling strategy based on the renewable energy power generation prediction results and system constraints, using a linear one-dimensional projection constraint sub-Blu-bar control method, and to solve the optimal scheduling scheme through mixed integer linear programming.

[0273] The distributed coordination and control module is used to achieve multi-device coordination and control based on the optimal scheduling scheme, using a push-pull average consensus algorithm with dynamic compression, achieving scheduling consensus through an adaptive quantization scheme, and performing real-time power adjustment operations.

[0274] The closed-loop feedback optimization module is used to continuously optimize the hybrid prediction model and scheduling strategy based on the actual execution status of the equipment and the operating status of the system, thereby forming an adaptive closed-loop control system.

[0275] This invention provides a method and system for renewable energy power generation prediction and power dispatch. By applying neural symbolic acceleration technology to renewable energy power prediction, physical constraints are encoded into symbolic rules and embedded into neural networks, significantly improving prediction accuracy and computational efficiency. A linear one-dimensional projection constraint-based distributed robust control method is employed to handle the uncertainty of renewable energy output. An uncertainty set is constructed using Wasserstein distance to ensure the robustness of the dispatch strategy under various scenarios. A push-pull averaging consensus algorithm with dynamic compression is introduced to achieve efficient coordination among distributed devices. An adaptive quantization scheme achieves optimal dispatch consensus without requiring global information. A unified optimization framework with multiple time scales, from minute-level prediction to daily-level dispatch, is constructed to organically combine short-term accurate prediction with long-term robust dispatch, improving the overall system operating efficiency.

[0276] This disclosure also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program performs the steps of renewable energy power generation prediction and power dispatch as described in the above method embodiments. The storage medium can be a volatile or non-volatile computer-readable storage medium.

[0277] In addition, this disclosure also provides a computer program product, which stores a computer program. When the computer program is run by a processor, it performs the steps of renewable energy power generation prediction and power dispatch provided in any of the above embodiments of this disclosure. For details, please refer to the above method embodiments, which will not be repeated here.

[0278] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium, which can be a volatile or non-volatile computer-readable storage medium. In another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0279] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices and apparatuses described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this disclosure, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces; the indirect coupling or communication connection of devices or units may be electrical, mechanical, or other forms.

[0280] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0281] In addition, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0282] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0283] Finally, it should be noted that the above-described embodiments are merely specific implementations of this disclosure, used to illustrate the technical solutions of this disclosure, and not to limit it. The protection scope of this disclosure is not limited thereto. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this disclosure; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be covered within the protection scope of this disclosure. Therefore, the protection scope of this disclosure should be determined by the protection scope of the claims.

Claims

1. A method of renewable energy power generation power prediction and power dispatching, characterized in that, The method comprises the following steps: Collecting historical power generation data and real-time meteorological data of renewable energy power generation, performing linear interpolation to fill in missing values and box plot anomaly detection on the historical power generation data and real-time meteorological data to obtain a normalized training data set; Using the normalized training data set, a hybrid prediction model is constructed by using a neural-symbolic acceleration technology with time logic constraints, the hybrid prediction model comprising an LSTM network and a CNN-Transformer network, short-term time sequence dependence is captured by the LSTM network and medium-long-term spatial and temporal features are extracted by the CNN-Transformer network, and a renewable energy power generation power prediction result is generated; According to the renewable energy power generation power prediction result and system constraint conditions, a linear one-dimensional projection constrained distribution robust control method is used to formulate a scheduling strategy, and the optimal scheduling scheme is solved by using the scheduling strategy through mixed integer linear programming, comprising: According to the renewable energy power generation power prediction result, a multi-objective scheduling model considering economy and safety is established, and a weighted sum method is used to unify the minimization of power purchase cost, the maximization of renewable energy consumption rate and the minimization of system frequency deviation into a single optimization target; Using the single optimization target, the complete system constraint conditions are constructed, including power balance constraint, device operation constraint, network security constraint and unit climbing constraint, to ensure the feasibility and safety of the scheduling scheme; Under the system constraint conditions, the linear one-dimensional projection constrained distribution robust control method is used to handle the uncertainty of renewable energy output, the difference between the real distribution and the empirical distribution is measured by using the Wasserstein distance, and the optimal scheduling scheme is obtained, comprising: based on the renewable energy power generation power prediction result, the confidence interval is defined as the center of the prediction value, the wind power uncertainty interval is set as 25% of the positive and negative prediction value, and the photovoltaic uncertainty interval is set as 20% of the positive and negative prediction value, to generate an initial uncertainty set; the difference between the real distribution and the empirical distribution of the initial uncertainty set is measured by using the Wasserstein distance, the Wasserstein radius is set as 0.1, the confidence level is determined as 95% based on the statistical characteristics of historical prediction errors, and the distribution robust uncertainty set is generated; in the uncertainty set, the distribution robust optimization problem is converted into a deterministic mixed integer linear programming problem by using the duality theory, the complexity of the probability constraint is eliminated, and the system safe operation requirement can still be met in the worst case, and the optimal scheduling scheme is obtained; Based on the optimal scheduling scheme, a push-pull average consensus algorithm with dynamic compression is used to realize multi-device coordinated control, a self-adaptive quantization scheme is used to achieve scheduling consensus, and real-time power regulation operation is performed; Based on the actual execution state of the device and the system operation state, the hybrid prediction model and the scheduling strategy are continuously optimized by using a reinforcement learning algorithm to form a self-adaptive closed-loop control system.

2. The method of claim 1, wherein, The historical power generation data and real-time meteorological data of renewable energy power generation are collected, the historical power generation data and real-time meteorological data are linearly interpolated to fill in missing values and box plot anomaly detection is performed, and a normalized training data set is obtained, including: The historical power generation data of wind power and photovoltaic power stations and real-time meteorological data of meteorological observation stations are collected through the SCADA system, the historical power generation data and real-time meteorological data are timestamp consistency checked, and synchronized original data are obtained; The quality of the synchronized original data is checked to identify data missing, outliers and duplicate records, linear interpolation is used to fill in missing values, outliers are identified by box plot method, and cleaned data are generated; The cleaned data are subjected to Min-Max normalization processing to map the data to the 0 to 1 interval, the periodic characteristics of hours, weeks and months are extracted, and lag features are generated, to obtain the normalized training data set.

3. The method of claim 2, wherein, Linear interpolation is used to fill in missing values, and outliers are identified by box plot method to generate cleaned data, including: The synchronized original data are subjected to missing value detection, when the continuous missing time is less than 2 hours, the linear interpolation method is used to calculate the missing value based on the data before and after the time, when the continuous missing time is more than 2 hours, the historical same period data is used to replace, and the completed data are generated; Based on the completed data, the first quartile, the third quartile and the interquartile range of each parameter are calculated, the abnormal data points exceeding 1.5 times the interquartile range are identified by box plot method, and the outlier marking result is generated; According to the outlier marking result, the abnormal data points are removed, the integrity of the data after removal is verified to ensure data continuity and consistency, and the cleaned data are obtained.

4. The method of claim 1, wherein, The LSTM network captures short-term time sequence dependence, including: Based on the normalized training data set, a multi-layer LSTM network is constructed, the hidden layer dimension is set to 128, and the information of key time steps is adaptively selected through attention mechanism; The power generation of the past 24 hours, meteorological data and equipment state information are input into the LSTM network for processing, short-term time sequence dependence features are extracted, and hourly power prediction values for the next six hours are generated; Time logic constraints are introduced to ensure that the predicted power change rate does not exceed the device climbing ability, and short-term power prediction results that meet physical constraints are obtained.

5. The method of claim 1, wherein, The CNN-Transformer network extracts medium and long-term spatial and temporal features, including: Based on the normalized training data set, a CNN-Transformer hybrid architecture is designed, where the CNN module uses one-dimensional convolution to extract local time sequence patterns, and the convolution kernel size is 3, 5 and 7 respectively; Multi-scale feature fusion is used to capture the change law of different periods, where the Transformer module uses a multi-head self-attention mechanism to handle long-term dependencies, the number of heads is 8, and the feedforward network dimension is 512; The neural symbolic acceleration technology is used to model the prediction problem as a constrained optimization problem, an approximate optimal solution is quickly solved by a neural network, and the result is verified and corrected by symbolic reasoning, so that the medium and long term space-time characteristics are obtained.

6. The method of claim 1, wherein, Based on the optimal scheduling scheme, a push-pull average consensus algorithm with dynamic compression is used to realize multi-device coordinated control, a self-adaptive quantization scheme is used to reach scheduling consensus, and real-time power regulation operations are performed, including: Based on the optimal scheduling scheme, the energy storage system, controllable load and distributed power in the power system are modeled as intelligent agents with autonomous decision-making capability, each agent is equipped with an independent computing unit and a communication module; A distributed protocol based on the push-pull average consensus algorithm is designed, each agent periodically exchanges information with a predefined neighbor agent set, and the neighbor relationship is determined based on the physical connection topology and the communication network structure; In the push operation phase, each agent sends the current state vector and gradient information to all neighbor agents, and the sending weight is dynamically adjusted according to the degree centrality of the network topology; In the pull operation phase, each agent receives state information from the neighbors, and updates the local state by fusing neighbor information through weighted averaging; A multi-level convergence criterion is established, and scheduling consensus is reached through state convergence, objective function convergence and consistency error criterion, and the real-time power regulation operation is performed.

7. The method of claim 6, wherein, The adaptive quantization scheme includes: A dynamic compression mechanism based on information theory is used to evaluate the importance of information according to Shannon entropy, and the quantization precision is adaptively adjusted to reduce the communication burden; A three-level quantization system is established, the key control signal is quantized with 16-bit high precision, the state monitoring information is quantized with 8-bit medium precision, and the auxiliary information is quantized with 4-bit low precision; The variance and change rate of the system operating state are monitored in real time, when the system operating state is in a steady state, a coarse quantization strategy is used to reduce communication overhead, when the system operating state is disturbed, the system automatically switches to a fine quantization mode to ensure control accuracy.

8. The method of claim 1, wherein, Based on the actual execution state of the device and the system operating state, the hybrid prediction model and the scheduling strategy are continuously optimized through a reinforcement learning algorithm to form an adaptive closed-loop control system, including: A real-time data feedback channel is established, the actual execution state of the device is collected through the SCADA system and the smart meter, including the actual charge and discharge power of the energy storage system, the actual adjustment amount of the controllable load, and the actual output of the renewable energy; A prediction error index is calculated, including mean absolute error, root mean square error and mean absolute percentage error, and when any index exceeds a preset threshold, a model retraining mechanism is triggered; An online learning algorithm is designed to update the prediction model parameters in an incremental learning manner, the sample weight of the updated sample is set to 0.1, and the historical sample weight is adjusted according to an exponential decay law; A scheduling performance evaluation system is constructed, including economic indicators, safety indicators and environmental protection indicators, and when the performance indicators deviate from the target value by more than 5%, a scheduling strategy adaptive adjustment mechanism is triggered. Based on the scheduling performance evaluation system, the scheduling strategy is continuously optimized by a reinforcement learning algorithm, the system operating state is taken as input, the scheduling action is taken as output, and a multi-objective weighted sum is taken as a reward function, forming the adaptive closed-loop control system.

9. The method of claim 8, wherein, The reinforcement learning algorithm comprises: Based on the scheduling performance evaluation system, the scheduling strategy is continuously optimized by a deep deterministic policy gradient algorithm, the system operating state is taken as input, the scheduling action is taken as output, and a multi-objective weighted sum is taken as a reward function; A state space and an action space are constructed, the state space comprises renewable energy prediction bias, energy storage state of charge, load change rate, and price information, and the action space comprises energy storage charging and discharging instructions, load adjustment instructions, and grid interaction power; An experience replay buffer is set to store historical state-action-reward sequences, the buffer size is 10000, batch learning is performed every 100 steps, and the batch size is 32; An actor network and a critic network are established, the actor network outputs continuous action values, and the critic network evaluates the value function of the state-action pair; The actor network parameters are updated by a gradient ascent method to maximize the cumulative reward, and the critic network parameters are updated by a time difference error, forming a continuously optimized reinforcement learning strategy.

10. The method of claim 1, wherein, A hybrid prediction model is constructed by using a neural-symbolic acceleration technique with time logic constraints, comprising: Based on the normalized training data set, the power change rate limit and the equipment capacity constraint are encoded as symbolic rules, and the constraint rules are embedded into the deep learning model through a neural-symbolic interface layer; A time logic constraint processing module is constructed to constrain the wind power change rate to be less than 10% of the rated power per minute and the photovoltaic power change rate to be less than 20% of the rated power per minute, and a constraint embedding vector is generated; The neural-symbolic acceleration technique of MILP motion planning is used to model the prediction problem as a constrained optimization problem, and the approximate optimal solution is quickly solved by a neural network, and the prediction result is verified and corrected by symbolic reasoning, to obtain the hybrid prediction model satisfying the time logic constraint.

11. A renewable energy power generation power prediction and power dispatch system, characterized by, Comprise: A data acquisition and preprocessing module is used to acquire historical power generation data and real-time weather data of renewable energy power generation based on a standardized data interface, and to perform linear interpolation to fill in missing values and box plot anomaly detection on the historical power generation data and real-time weather data, to obtain a normalized training data set; A neural-symbolic fusion prediction module is used to construct a hybrid prediction model based on the normalized training data set by using a neural-symbolic acceleration technique with time logic constraints, to capture short-term time series dependencies by an LSTM network and to extract medium and long-term spatial and temporal features by a CNN-Transformer network, and to generate renewable energy power generation prediction results; A distributed robust scheduling optimization module is used to develop a scheduling strategy based on the renewable energy power generation prediction results and system constraint conditions by using a linear one-dimensional projection constrained distributed robust control method, and to solve an optimal scheduling scheme by a mixed integer linear programming, comprising: According to the renewable energy power generation prediction result, a multi-objective scheduling model considering economy and safety is established, and a weighted sum method is used to unify the minimization of power purchase cost, the maximization of renewable energy consumption rate and the minimization of system frequency deviation into a single optimization objective; The single optimization objective is used to build complete system constraint conditions, including power balance constraint, equipment operation constraint, network safety constraint and unit climbing constraint, to ensure the feasibility and safety of the scheduling scheme; Under the system constraint conditions, a linear one-dimensional projection constraint distribution robust control method is used to process the uncertainty of renewable energy output, a Wasserstein distance is used to measure the difference between the real distribution and the empirical distribution, and the optimal scheduling scheme is obtained, including: based on the renewable energy power generation prediction result, a confidence interval is defined with the predicted value as the center, a wind power uncertainty interval is set as plus or minus 25% of the predicted value, and a photovoltaic uncertainty interval is set as plus or minus 20% of the predicted value, and an initial uncertainty set is generated; the difference between the real distribution and the empirical distribution of the initial uncertainty set is measured by the Wasserstein distance, the Wasserstein radius is set as 0.1, the confidence level is determined as 95% based on the statistical characteristics of historical prediction errors, and the distribution robust uncertainty set is generated; in the uncertainty set, the dual theory is used to convert the distribution robust optimization problem into a deterministic mixed integer linear programming problem, the complexity of the probability constraint is eliminated, and it is ensured that the system safety operation requirement can still be met in the worst case, and the optimal scheduling scheme is obtained; A distributed coordination control module is used to realize multi-device coordination control based on the optimal scheduling scheme by using a push-pull average consensus algorithm with dynamic compression, to achieve scheduling consensus through an adaptive quantization scheme, and to perform real-time power regulation operation; A closed-loop feedback optimization module is used to continuously optimize the hybrid prediction model and scheduling strategy based on the actual execution state of the device and the system operation state through a reinforcement learning algorithm, to form an adaptive closed-loop control system.

Images