An AI-based intelligent park source-grid-load-storage-charging integrated scheduling method

Abstract

The present invention discloses an AI-based intelligent park source-grid-load-storage-charging integrated scheduling method, which relates to the field of energy scheduling technology. The method includes real-time collection and preprocessing of park equipment data to generate a standardized time series data matrix; using time series information to filter and extract data features of the standardized time series data matrix, and using a hybrid neural network framework to construct a source-grid-load-storage-charging optimization model to obtain an embedded vector; using game theory to optimize the preliminary scheduling strategy and generate a collaborative scheduling instruction set; using protocol conversion to convert the collaborative scheduling instruction set into device control instructions and execute them; using dynamic threshold filtering to monitor the device operating status, safety boundaries, and economic thresholds in real time, and using online learning to optimize device control instructions. The present invention improves the economic efficiency of the scheduling strategy by using a proximal strategy optimization pruning algorithm and a Nash equilibrium game.

Description

Technical Field

[0001] The present invention relates to the field of energy scheduling technology, and in particular to an AI-based intelligent park source-grid-load-storage-charging integrated scheduling method. Background Art

[0002] With the rapid transformation of energy structures and the development of smart grids, energy management in smart parks is gradually evolving towards integrated power generation, grid-loading, storage, and charging. Current smart park scheduling methods mostly utilize traditional optimization algorithms and rule-based strategies, further enabling coordinated control of energy devices through centralized or hierarchical architectures. These methods generally employ fixed thresholds or static models to make scheduling decisions, combined with forecasting techniques to provide short-term load and power generation forecasts. Furthermore, most existing technologies utilize standardized communication protocols to enable data exchange between devices, and secure encryption technology is used to ensure the reliability of command transmission. With the rapid development of the Energy Internet and distributed energy, this provides fundamental support for the efficient operation of parks.

[0003] However, existing methods still lack dynamic collaborative optimization and real-time adaptability. Traditional optimization algorithms cannot effectively handle the complex game relationships between multiple devices and multiple objectives, which means that the global optimality of the scheduling strategy is insufficient. At the same time, the use of static models and fixed threshold mechanisms cannot fully adapt to the rapid changes in the park's operating status. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an AI-based smart park source-grid-load-storage-charging integrated scheduling method to solve the problems of deficiencies in dynamic collaborative optimization and real-time adaptability of smart parks.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In the first aspect, the present invention provides an AI-based intelligent park source-grid-load-storage-charging integrated scheduling method, which includes real-time collection of park equipment data and preprocessing to generate a standardized time series data matrix; using time series information to filter and extract data features of the standardized time series data matrix, and using a hybrid neural network framework to construct a source-grid-load-storage-charging optimization model to obtain an embedding vector; mapping the embedding vector to the action space through a policy network, and using a proximal policy optimization and clipping algorithm to obtain a preliminary scheduling strategy; using game theory to optimize the preliminary scheduling strategy to generate a collaborative scheduling instruction set; using protocol conversion to convert the collaborative scheduling instruction set into equipment control instructions and execute them; using dynamic threshold filtering to monitor the equipment operating status, safety boundaries and economic thresholds in real time, and using online learning to optimize equipment control instructions.

[0008] As a preferred solution of the AI-based intelligent park source-grid-load-storage-charging integrated scheduling method of the present invention, wherein: the real-time collection of park equipment data and pre-processing to generate a standardized time series data matrix, the specific steps are as follows:

[0009] Time synchronization is performed through the time protocol, and the sampling frequency is adjusted using the dynamic sampling rate to collect data from campus equipment;

[0010] Based on the park equipment data, sliding window smoothing and denoising are adopted, and dynamic Z score is used for standardization to construct standardized park equipment data;

[0011] The standardized campus equipment data is cleaned using linear interpolation completion and anomaly detection, and the cleaned campus equipment data is converted into a standardized time series data matrix through time series alignment.

[0012] As a preferred solution of the AI-based intelligent park source-grid-load-storage-charging integrated scheduling method of the present invention, wherein: the data features of the standardized time series data matrix are extracted by using time series information screening, and a hybrid neural network framework is used to construct a source-grid-load-storage-charging optimization model to obtain an embedded vector. The specific steps are as follows:

[0013] A hybrid neural network architecture is used to combine spatiotemporal encoders, feature extractors, and decision makers to build a source-grid-load-storage optimization model.

[0014] The spatiotemporal encoder uses a bidirectional long short-term memory network to perform spatiotemporal position encoding, converting the standardized time series data matrix into an enhanced time series feature matrix;

[0015] The feature extractor is based on the enhanced temporal feature matrix, combined with Hilbert space projection and causal entanglement gating, to extract quantum attention features in the quantum state space;

[0016] Based on the quantum attention features, the device association matrix is ​​generated by quantum probability statistics method, and the embedding vector is obtained in combination with the decision maker.

[0017] As a preferred solution of the AI-based intelligent park source-grid-load-storage-charging integrated scheduling method of the present invention, wherein: the embedding vector is mapped to the action space through the policy network, and the proximal policy optimization and clipping algorithm is used to obtain the preliminary scheduling strategy. The specific steps are as follows:

[0018] A three-branch fully connected neural network is used to map the embedding vector to the initial action space and generate the initial action vector;

[0019] Use masked temperature Softmax combined with device physical constraints to obtain the feasible strategy probability of the initial action vector;

[0020] Based on the feasible strategy probability, the proximal strategy optimization pruning method and L2 projection correction are combined to optimize the feasible strategy probability, and the Lagrange multiplier method is used to generate the preliminary scheduling strategy.

[0021] As a preferred solution of the AI-based intelligent park source-grid-load-storage-charging integrated scheduling method of the present invention, wherein: the preliminary scheduling strategy is optimized using game theory to generate a collaborative scheduling instruction set. The specific steps are as follows:

[0022] Use game theory to adjust and optimize the initial scheduling strategy, and use multi-objective optimization to screen the non-inferior solution set and generate the optimized adjustment strategy;

[0023] Using semantic templates, the optimization adjustment strategy is mapped into device control parameters and a collaborative scheduling instruction set is generated.

[0024] As a preferred solution of the AI-based intelligent park source-grid-load-storage-charging integrated scheduling method of the present invention, wherein: the protocol conversion is used to convert the collaborative scheduling instruction set into device control instructions and execute them. The specific steps are as follows:

[0025] Use structured data parsing to parse the collaborative scheduling instruction set, extract the device number, control parameters and execution time, and generate device control instructions;

[0026] Use industrial protocol converter to convert device control instructions into communication protocol format;

[0027] Use data encryption transmission to send and execute device control instructions through a secure channel.

[0028] As a preferred solution of the AI-based intelligent park source-grid-load-storage-charging integrated scheduling method described in the present invention, the method utilizes dynamic threshold filtering to monitor the equipment operating status, safety boundary, and economic threshold in real time, and uses online learning to optimize equipment control instructions. The specific steps are as follows:

[0029] Use edge computing nodes to collect real-time campus equipment data and combine it with preprocessing to generate a standardized data matrix;

[0030] Through standardized data matrix, real-time working condition analysis is used to obtain safety boundary thresholds and economic thresholds, and a dynamic threshold table is constructed;

[0031] Based on the dynamic threshold table, federated learning is used to optimize device control instructions, perform quantum verification, and generate a verification report.

[0032] Adopt hierarchical execution strategy and feedback optimization to generate optimized equipment control instructions.

[0033] As a preferred solution of the AI-based intelligent park source-grid-load-storage-charging integrated scheduling method of the present invention, wherein: the use of federated learning to optimize device control instructions, and perform quantum verification to generate a verification report, the specific steps are as follows:

[0034] Based on the dynamic threshold table, distributed strategy optimization is used to preliminarily optimize device control instructions;

[0035] According to the optimized equipment control instructions, multimodal simulation verification is used to perform quantum verification and digital twin testing, and a verification report is generated.

[0036] In a second aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the AI-based intelligent park source-grid-load-storage-charging integrated scheduling method as described in the first aspect of the present invention.

[0037] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the AI-based smart park source-grid-load-storage-charging integrated scheduling method as described in the first aspect of the present invention.

[0038] The beneficial effects of the present invention are: through the hybrid neural network architecture combined with quantum state space feature extraction, the dynamic correlation characteristics between devices are deeply mined, and the characterization ability of spatiotemporal features is improved; the proximal strategy is used to optimize the clipping algorithm and Nash equilibrium game to improve the economy of the scheduling strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1 This is a flow chart of the AI-based smart park integrated source-grid-load-storage-charging scheduling method.

[0041] Figure 2 Flowchart of data collection and preprocessing.

[0042] Figure 3 Flowchart of the hybrid neural network architecture.

[0043] Figure 4 It is the instruction execution flow chart. DETAILED DESCRIPTION

[0044] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0045] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0046] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0047] Reference Figures 1 to 4 , is an embodiment of the present invention, which provides an AI-based intelligent park source-grid-load-storage-charging integrated scheduling method, including the following steps:

[0048] S1. Collect and pre-process park equipment data in real time to generate a standardized time series data matrix.

[0049] Furthermore, time synchronization is performed through a time protocol, and the sampling frequency is adjusted using a dynamic sampling rate to collect data from park equipment.

[0050] Specifically, the precise time protocol is used to calibrate the local clocks of campus equipment for time alignment, reducing timing misalignment caused by network delays and device clock drift. The sampling rate is dynamically adjusted, with a fixed high sampling rate (1kHz) for fast-changing signals such as current and voltage, and a fixed low sampling rate (1Hz) for slow-changing signals such as temperature and SOC, to collect data from campus equipment.

[0051] It should be noted that park equipment data includes energy equipment data, load data, environment and operating status data, time and topology data, and economic and strategy data. Energy equipment data refers to the park load demand and historical load, load data refers to the park load demand and historical load, environment and operating status data refers to environmental parameters (temperature, humidity, light intensity and weather forecast data), equipment operating status and safety boundary parameters (such as grid frequency fluctuation limit and equipment power upper limit), time and topology data refers to time synchronization information and physical topology data, and economic and strategy data refers to cost parameters and optimization goals (economic, environmental and fairness).

[0052] Based on the park equipment data, sliding window smoothing and denoising are adopted, and dynamic Z score is used for standardization to construct standardized park equipment data.

[0053] Specifically, a sliding window is used for smoothing and denoising. A narrow window is used for fast-changing signals such as current and power, and a wide window is used for smooth signals such as temperature. The arithmetic mean of the multidimensional time series data stream is obtained within the sliding window, high-frequency noise is suppressed, the mean and standard deviation of the data within the window are obtained, and each data point is standardized using a dynamic Z score to construct standardized initial data.

[0054] The standardized campus equipment data is cleaned using linear interpolation completion and anomaly detection, and the cleaned campus equipment data is converted into a standardized time series data matrix through time series alignment.

[0055] Specifically, based on the standardized initial data, linear interpolation is used to complete the missing data points using linear interpolation of the two valid data points before and after. The 3σ principle is used for anomaly detection to obtain the mean μ and standard deviation σ, and the data points exceeding μ±3σ are eliminated. For data missing caused by communication interruption, linear interpolation is used to generate estimated values ​​for data completion. The time series alignment method is used to resample all campus equipment data based on a fixed time interval (such as 1 second), and the timestamps are strictly aligned. Each row represents a campus device, and each column represents a time point. The corresponding standardized data values ​​are filled in to generate a standardized time series data matrix.

[0056] It should be noted that the standardized time series data matrix contains the time steps and the number of variables, such as photovoltaic power, load demand, energy storage status, etc.

[0057] Preferably, the precise time protocol is used to calibrate the device clock, reduce network latency and clock drift, and improve the accuracy and synchronization of data collection; sliding window filtering is used for denoising and abnormal data processing to enhance data quality and anti-interference capabilities; and Z-value normalization is used to eliminate unit differences between different parameters.

[0058] S2. Use time series information to filter and extract the data features of the standardized time series data matrix, and use a hybrid neural network framework to build a source-grid-load-storage-charging optimization model to obtain the embedding vector.

[0059] Furthermore, a hybrid neural network architecture is used to combine spatiotemporal encoder, feature extractor and decision maker to construct a source-grid-load-storage-charging optimization model.

[0060] Specifically, the construction of the spatiotemporal encoder adopts a graph convolutional network, which takes the physical connection topology of the park energy equipment (such as the distribution network bus structure, photovoltaic array layout) as input, constructs an adjacency matrix (edge ​​weights may include line impedance or communication delay), aggregates neighbor node information through multi-layer graph convolution operations, and superimposes a time convolution network on the basis of spatial features. The multi-scale receptive field of the expanded convolution is used to capture long-period timing patterns, and realizes the joint encoding of physical topology and timing dynamics; a multi-head causal attention mechanism is adopted to apply a masked attention layer to the spatiotemporal feature tensor. The mask matrix is ​​generated by the device communication delay and causal timing constraints to obtain the dynamic correlation weights between devices (such as the impact of photovoltaic output fluctuations on the energy storage charging and discharging priority), combined with the residual connection to retain the initial The initial features are used to alleviate the gradient vanishing problem, and the training process is stabilized through layer normalization, which includes explicit and implicit features of device interaction. The decision maker's optimization strategy uses a policy-value dual-branch structure. The policy branch's fully connected layer outputs the action probability distribution (such as the continuous value interval of the energy storage charging and discharging power), and directly uses the hyperbolic tangent function to constrain the output range. The value branch evaluates the state value (such as the expected cost reduction) by the evaluation network, guides the policy gradient update, and embeds physical constraints. Lagrange multipliers are used in the output layer to convert the device safety boundaries (such as SOC limits and PCS power caps) into differentiable constraints, and directly optimize the constrained strategy. The spatiotemporal encoder, feature extractor and decision maker are cascaded in an end-to-end manner to construct a source-grid-load-storage-charging optimization model.

[0061] It should be noted that the physical constraints include the physical constraints of energy storage equipment, the physical constraints of grid interaction, the physical constraints of load regulation, and other equipment constraints. The physical constraints of energy storage equipment include SOC limitations, charge and discharge mutual exclusivity, and power limitations. The physical constraints of grid interaction include forced switching between grid and off-grid and power exchange limitations. The physical constraints of load regulation include load protection and adjustable range limitations. Other equipment constraints include photovoltaic inverter output limitations and dynamic adjustment of charging pile power. The strategy value dual-branch structure will evaluate the value of the current state, generate the strategy value based on the extracted features, and use the strategy branch fully connected network to output the action probability distribution. This probability distribution reflects the expected benefits of taking different actions under different states. The strategy value is obtained based on the expected cost reduction.

[0062] The spatiotemporal encoder uses a bidirectional long short-term memory network for spatiotemporal position encoding, converting the standardized time series data matrix into an enhanced time series feature matrix.

[0063] Specifically, based on the standardized time series data matrix, a bidirectional long short-term memory network and spatiotemporal position encoding are used. The forward time series analysis analyzes the data in the standardized time series data matrix point by point from the first time point to the last time point to capture the historical change pattern (for example, analyzing how the photovoltaic power generation power is affected by the light intensity in the previous hour). The reverse time series analysis is reversed from the last time point to the first time point to predict possible future change trends (for example, adjusting the energy storage charging strategy in advance based on the weather forecast for the next two hours). The forward and reverse analysis results are merged to form a bidirectional time series feature that can reflect historical laws and predict future trends; according to the actual location of the equipment in the park (such as the location of the photovoltaic panel), the The system uses the latitude and longitude coordinates and the installation location of the energy storage equipment to obtain the spatial relationship of the park equipment (such as how far it is from the nearest substation and whether it is in a shadowed area). Combined with time information (such as whether the current electricity price is peak or trough, and the impact of seasonal changes on power generation), different weights are assigned to the park equipment data in different time periods. The location information and time information of the equipment are combined to construct a time and position code. The bidirectional time series features and spatiotemporal position codes are aligned and superimposed at the time points, retaining features (such as sudden load changes or extreme weather impacts) and filtering out irrelevant noise. At the same time, short-term fluctuations (such as minute-level power changes) and long-term trends (such as seasonal power consumption peaks) are captured, and an enhanced time series feature matrix is ​​output.

[0064] The feature extractor is based on the enhanced temporal feature matrix, combined with Hilbert space projection and causal entanglement gating, to extract quantum attention features in the quantum state space.

[0065] Specifically, based on the enhanced timing feature matrix, combined with Hilbert space projection and causal entanglement gating, each device state in the enhanced timing feature matrix (such as photovoltaic output, energy storage SOC) is converted into a quantum state representation, and the characteristics of the quantum state (the superposition state of quantum bits) are used to measure the probability of high and low current photovoltaic power generation. The timing periodicity is encoded into phase information through quantum Fourier transform, and the Hamiltonian operator is obtained through energy eigenstate mapping according to the device physical equations such as the energy storage charging and discharging power equation. The features are projected into the energy eigenstate space, and the numerical features are converted into quantum states that represent probability and correlation; controlled NOT gates are used to connect devices with strong coupling, and the control bit is controlled by the photovoltaic output fluctuation. The mark is composed of energy storage charge and discharge instructions, and the state such as grid frequency deviation is transmitted through the quantum teleportation protocol. Based on the real-time interaction data of the devices such as the power flow direction, the von Neumann entropy is obtained, the entanglement strength is quantified, and the quantum annealing algorithm is used to optimize the entangled topology to construct an entangled state network; using quantum measurement and classical mapping, the entangled state network is measured by Pauli Z basis, the expected value of each quantum bit is extracted, and the dimension is reduced by quantum principal component analysis. The top k quantum features with the largest variance are retained. According to the measurement results, the instantaneous mutual information between devices is obtained (such as the mutual information between photovoltaics and energy storage reflects their synergistic efficiency), and a dynamic adjacency matrix is ​​constructed. The matrix element value represents the causal strength, ranging from 0 to 1, and the quantum attention feature is extracted.

[0066] Based on the quantum attention features, the device association matrix is ​​generated by quantum probability statistics method, and the embedding vector is obtained in combination with the decision maker.

[0067] Specifically, based on the quantum state probability value of each device in the quantum attention feature, such as the probability of high power generation of photovoltaic equipment of 0.7, it is converted into a probability value through Pauli Z basis measurement, and a sliding time window (5 consecutive time points) is used to count the frequency of occurrence of device state combinations to generate joint probability (such as the probability that the photovoltaic output is greater than 80% and the energy storage is charged) and edge probability (probability of a single device state); based on the joint probability and edge probability, the information entropy formula is used to obtain the information entropy of all devices and device pairs, and the information entropy formula is used to directly obtain the joint strength value, perform symmetry processing, take the average value of the two-way correlation and normalize it to the range of [0,1], and construct a device association matrix; the device association matrix is ​​input into the decision maker for processing, and the attention mechanism in the decision maker is used to calculate the device importance weight. The feature is made reasonable through the constraints of the decision maker, and the embedding vector is output.

[0068] It should be noted that the constraints are to limit the characteristic values ​​within the allowable range, such as 20%-90% of the energy storage charge, to verify the energy balance relationship and the correction of abnormal characteristics; the quantum state probability value is the device state (such as photovoltaic power generation, energy storage charge and discharge status) represented by quantum bits, and the Pauli Z-basis measurement projects the quantum state to probability, and directly takes the square modulus of the probability amplitude when actually acquiring each quantum state; the embedded vector contains the comprehensive timing characteristics of equipment such as power supply, power grid, load, energy storage, and charging piles in the park.

[0069] The better approach is to ensure temporal causality through a masking mechanism and improve the credibility of the scheduling strategy; use bidirectional LSTM spatiotemporal coding to forward analyze historical patterns and reversely predict future trends, enhancing robustness to load fluctuations and weather changes; and combine the physical location of the equipment to generate spatiotemporal location coding to improve spatial correlation.

[0070] S3. Map the embedding vector to the action space through the policy network, and use the proximal policy optimization pruning algorithm to obtain the preliminary scheduling strategy.

[0071] A three-branch fully connected neural network is used to map the embedding vector to the initial action space and generate the initial action vector.

[0072] Specifically, the embedding vector is input into the three-branch fully connected neural network layer. The function of the fully connected neural network layer is to extract the common features of all devices, such as electricity price fluctuations, weather impacts, overall load trends, etc., to form shared features. The shared features are input into three independent neural network branches respectively, and each branch is responsible for generating control actions for different devices; Branch 1 energy storage and charge and discharge control, the structure is a multi-layer fully connected network that ultimately outputs two values, representing charging power and discharging power respectively. The constraint is that charging and discharging will not occur at the same time, such as prohibiting charging when the SOC is full. The output is the charging and discharging instructions of the energy storage device, such as charging 300kW or discharging 200kW; Branch 2 power grid Interactive decision-making is structured as a multi-layer fully connected network, and ultimately outputs a probability value between 0 and 1, indicating whether to operate in parallel with the grid. The constraint is that if a grid fault is detected, an off-grid instruction is forcibly output, and the output is a decision on whether to connect to the grid or go off-grid. Branch 3 load regulation is structured as a multi-layer fully connected network, and ultimately outputs priority scores for multiple loads. The constraint is that loads such as data centers always have the highest priority and cannot be disconnected. The output is a list of loads that need to be adjusted, such as reducing the air conditioning load by 20% and delaying production tasks. The outputs of the three branches are merged, including energy storage charging and discharging instructions, grid interactive decision-making, load regulation plans, etc., and normalized to generate an initial action vector.

[0073] The masked temperature Softmax is used in combination with the physical constraints of the device to obtain the feasible strategy probability of the initial action vector.

[0074] Specifically, the physical constraints of the equipment are set as the energy storage constraint SOC upper limit (overcharging is prohibited), charge and discharge mutual exclusion and maximum power limit, the grid constraint is the forced off-grid and grid-connected power limit under fault conditions, the load constraint is the key load protection and the maximum adjustable range, and the constraint compliance check is performed on each action suggestion in the initial action vector. For example, if the energy storage SOC=100%, the charging power suggestion will automatically fail, and if the grid fault signal is true, the grid-connected suggestion will automatically fail. A binary mask is used to generate a 0 or 1 mask value for each action dimension, where 0 indicates that the constraint must be prohibited and 1 indicates compliance. Constraints are allowed to be considered, and the mask value is updated according to the real-time device status. For example, when the energy storage SOC rises to 95%, the charging power mask value changes from 1 to 0.5 (soft limit). When there is a load marked as a temporary load, the temporary load adjustment mask value is immediately set to 0; the initial temperature is set according to the scheduling stage, the temperature is lowered in the early stage of training, the temperature is lowered when the voltage fluctuates greatly, and the temperature is increased when the load is high to promote exploration at high temperature. The operation stage is low temperature to enhance determinism and combined with real-time dynamic adjustments, such as when the SOC approaches the limit; and each action dimension is acquired and normalized to generate a feasible strategy probability.

[0075] It should be noted that the feasible strategy probability is the probability distribution of legal actions retained after physical constraint verification and masking, so that the scheduling strategy complies with the equipment operation restrictions, such as the energy storage SOC limit and the power exchange upper limit of the grid, and avoids invalid and dangerous instructions.

[0076] It should be noted that the formula for calculating the probability of a feasible strategy is: ;

[0077] in, is the probability of a feasible strategy, , is a binary mask, is 0 (disable) or 1 (allow), The strategy value includes temperature normalization, represents normalization, summing only the actions, represents the base of natural logarithm, Indicates strategic value An indexation of the attractiveness of device actions.

[0078] Based on the feasible strategy probability, the proximal strategy optimization pruning method and L2 projection correction are combined to optimize the feasible strategy probability, and the Lagrange multiplier method is used to generate the preliminary scheduling strategy.

[0079] Specifically, the proximal strategy is optimized to clip the update amplitude control, set the strategy update trust region such as the ±20% variation range, obtain the probability ratio of the new and old strategies (new strategy probability / old strategy probability), and perform clipping when the probability ratio of the new and old strategies exceeds the set range. The generalized advantage estimation method is used to set a discount factor based on immediate benefits and long-term impacts. The discount factor can be set to 0.9, and L2 projection correction is used to establish equipment operation constraints such as energy storage SOC range, charging and discharging power limits and minimum operating time. The projection correction is used to optimize the constrained quadratic programming, directly obtain the nearest feasible point and solve it using the Lagrange multiplier method; the Lagrange multiplier method is used to assign the strategy optimization tasks of different devices to multiple computing units, and an asynchronous update mechanism is used to improve efficiency. A calculation timeout protection such as a 500ms time limit is set, a small regularization term is added to prevent matrix singularity, gradient clipping is used to prevent explosion and perform real-time numerical overflow detection; the learning rate is dynamically adjusted according to the convergence situation, and optimization progress monitoring indicators are set to generate a preliminary scheduling strategy.

[0080] It should be noted that the default value for the probability ratio of the new and old strategies is 0.2. This ratio decreases when safety requirements are high (0.1-0.15) and increases when strong exploration is required (0.25-0.3). For example, it increases by 0.05-0.1 during the trial run phase, decreases by 0.05-0.1 during extreme weather warnings, and remains at the default value during peak electricity price periods. The learning rate is dynamically adjusted to 1.05% after three consecutive iterations of profit improvement. The learning rate remains unchanged when the profit rises and falls, and is 70% after two consecutive profit declines. The optimization progress monitoring indicators are set at the beginning of the optimization (taking the average of the first three rounds) and three indicators are updated after each iteration. When a color alert is triggered, yellow indicates logging and continued operation, and red indicates the implementation of emergency measures. The color alerts are green (all indicators are within the normal range), yellow (one indicator exceeds the range but does not reach the threshold (probability ratio), and red (two or three indicators exceed the threshold or a single indicator is seriously abnormal). The proximal policy optimization clipping method is used to map the embedding vector to the action space and generate a preliminary scheduling policy that conforms to the physical constraints of the equipment, maximizing the scheduling benefit while ensuring policy stability.

[0081] Preferably, a masking mechanism is used to force the probability of actions that violate physical constraints to zero, so that the generation of dangerous instructions is zero. The Lagrange multiplier hard constraint converts the device safety boundary into a differentiable loss term, so that the strategy meets physical constraints and enhances the safety of scheduling. A three-branch strategy network is used with dynamic adjustment of temperature parameters to enhance the rationality of the action.

[0082] S4. Use game theory to optimize the preliminary scheduling strategy and generate a collaborative scheduling instruction set.

[0083] Game theory is used to adjust and optimize the preliminary scheduling strategy, and multi-objective optimization is used to screen the non-inferior solution set to generate the optimized adjustment strategy.

[0084] Specifically, the contribution value allocation method in cooperative game theory is used to analyze the degree of influence of each device in the overall scheduling strategy. On the basis of the preliminary scheduling plan, the impact of each device on the total cost, grid stability and other indicators after adjustment is simulated to obtain the actual contribution ratio of each device. For example, the discharge of energy storage during peak electricity price periods can reduce the cost of purchasing electricity, and the contribution value will increase accordingly. The distribution fairness index in economics is used to evaluate the contribution distribution between devices in combination with pre-set constraints. First, the distribution difference of the contribution value of each device is obtained, and then it is checked whether the minimum contribution ratio requirements of each type of equipment are met. When it is found that the contribution value of the device is long-term low, optimization suggestions are generated, such as adjusting the scheduling priority of the charging pile to increase its participation. The distribution difference and the actual contribution ratio are integrated into a structured evaluation result, including the specific contribution value of each device. , overall fairness and targeted optimization suggestions, and generate scheduling strategy evaluation results; regard the park equipment as game participants, and the strategy space of each game participant is an adjustable power range such as energy storage charging and discharging power ±100kW. Define the profit function of each device, photovoltaic to maximize the absorption rate (reduce abandoned light), energy storage to maximize the peak-valley price difference profit, charging pile to minimize the charging waiting time, and output game parameters; based on the game parameters, map the Nash equilibrium solution set to the target space (cost, carbon emissions, fairness), generate the Pareto frontier, use the ε-constraint method to take the main goal such as cost minimization as the benchmark, impose constraints on the secondary goal (such as carbon emissions), screen non-inferior solutions, such as selecting the solution with the lowest carbon emissions under the constraint that the cost increase is less than or equal to 5%, output the Pareto optimal solution (1 to 3 groups of recommended scheduling strategies), and output the screening optimization adjustment strategy.

[0085] Using semantic templates, the optimization adjustment strategy is mapped into device control parameters and a collaborative scheduling instruction set is generated.

[0086] Specifically, semantic template matching is used to extract numerical values ​​from optimization strategies based on control parameters such as device types such as photovoltaics, energy storage, and charging piles (e.g., energy storage discharge power of 50kW and photovoltaic output limit of 80%). Contextual association analysis is then used to verify the feasibility of strategy parameters (e.g., preventing forced discharge when the SOC is less than 20%) in combination with the real-time status of the device (e.g., energy storage SOC of 60%). A structured control parameter table containing the device ID, parameter name, target value, and execution time window is then output. An industrial protocol converter is used to convert parameter values ​​into an instruction format supported by the device. Instructions with timing dependencies (e.g., locking the photovoltaic inverter before adjusting the energy storage power) are logically sorted to generate a control instruction set. Boundary value checking is used to verify whether the control instruction set exceeds the limit (e.g., the charging pile power instruction is less than or equal to the rated value). The SM2 algorithm is used to generate a to-be-executed instruction set from the instruction signature. Transmission priority is set based on the urgency of the instruction (e.g., fault handling > economic dispatch), and the instruction is sent to the device through protocols such as real-time instructions and non-real-time configuration to generate a collaborative scheduling instruction set.

[0087] The better approach is to use cooperative game theory and dynamic association matrix to enhance the collaboration of multiple devices; the dynamic adjacency matrix is ​​used to display the interactions between devices, providing quantitative indicators of the Pareto solution set, making decisions transparent and improving the explainability of strategies.

[0088] S5. Use protocol conversion to convert the collaborative scheduling instruction set into device control instructions and execute them.

[0089] Use structured data parsing to parse the collaborative scheduling instruction set, extract the device number, control parameters and execution time, and generate device control instructions.

[0090] Specifically, based on the collaborative scheduling instruction set, the collaborative scheduling instruction set is analyzed through the pattern matching method to determine whether it meets the format requirements. If a format abnormality is found, such as missing fields or data errors, the default value is automatically filled in based on historical data. If it cannot be filled in, it is marked as an abnormal format, and the device information library is queried to verify whether the device number is legal. The collaborative scheduling instruction set is only sent to registered devices, and standardized instructions are generated that only retain information such as device numbers, control parameters and execution time; control parameters such as power, voltage, current, etc. are disassembled from standardized instructions, and parameter names are converted into internal terms through semantic mapping, such as mapping power to active power, mapping voltage to line voltage, and converting non-standard units into internal benchmark units of the device, such as converting kilowatt-hours to joules, and generating a semantic standardized parameter table; adopting Using the time synchronization protocol and sliding time window method, the time information in the semantic standardized parameter table is converted into the internal time format of the device. The local time of the device is compared with the standard time server. If a deviation is found, time compensation is automatically performed. According to the scheduling cycle, such as 15 minutes, the standardized instruction execution time is adjusted to the starting point of the nearest scheduling window for calibration; templated instruction generation and physical constraint verification are used. The instruction template is called according to the device type, and the standardized parameters are filled in to generate the initial control instruction. For example, the charging and discharging instruction template of the energy storage device sets the active power to n kilowatts and the duration to m seconds. Check whether the control parameters exceed the device capability range, such as whether the power exceeds the maximum value. If it is found to be out of limit, it is automatically adjusted to a safe range, and the content of the initial control instruction is digitally signed to generate the device control instruction.

[0091] It should be noted that physical constraint verification refers to the real-time compliance check of instruction parameters when converting the generated collaborative scheduling instruction set into specific device control instructions, so that the instruction parameters comply with the physical operating limitations and safety boundaries of the device; the device information library is used to store and manage static and dynamic information of all connected devices, and is dynamically constructed by combining manufacturer data, manual configuration and real-time learning.

[0092] An industrial protocol converter is used to convert device control instructions into communication protocol format.

[0093] Specifically, it receives device control instructions, extracts the communication protocol of the target device from the device metadata database, analyzes the instruction format, data encoding method and communication rules supported by the communication protocol, and uses instruction semantic decomposition to decompose the device control instructions into operation units that can be recognized by the underlying protocol, such as setting the active power to 50kW into register address, writing numerical values ​​and other operations; matches the most suitable protocol template for the target device in the protocol knowledge base, combines the protocol version, device model and communication interface and other features, establishes a field mapping relationship according to the protocol specification, generates conversion rules including data format, byte order and check method, and performs preprocessing according to the special requirements of the protocol, such as the Modbus protocol needs to split 32-bit floating point numbers into two 16-bit registers, and generates the optimal protocol conversion solution; based on the optimal protocol conversion solution, converts the standardized parameters into the data format specified by the protocol, and assembles the communication message according to the protocol specification, adds the protocol header, address field, function code, data field and check code, etc., and adds timing control information according to the protocol requirements and converts it into the communication protocol format.

[0094] Based on the communication protocol format, data encryption is used for transmission, and device control instructions are sent and executed through a secure channel.

[0095] Specifically, based on the communication protocol format, protocol feature matching is used to analyze the communication protocol features of the target device, extract fields such as the transaction identifier and protocol identifier, and insert a security identifier into the protocol reserved field, keeping the original protocol frame structure unchanged. Encryption processing is implemented in the data domain for protocol compatibility encapsulation. The SM2 algorithm is used for key negotiation to generate a temporary session key (valid for 5 minutes). Layered encryption processing is adopted, control parameters are encrypted using SM4-CBC mode, auxiliary information is encrypted using lightweight encryption, and the SM3 algorithm is used to generate a message authentication code. A multi-mode secure transmission channel is used, and the transmission channel is selected according to the device security level. Key devices use a quantum key distribution channel, and ordinary devices use an IPSec VPN tunnel. Real-time channel monitoring is performed, continuously measuring channel quality indicators (latency, packet loss rate), dynamically adjusting the encryption strength (such as reducing it from 256 bits to 128 bits), and automatically switching to a backup channel when the main channel is interrupted. Decryption is completed within the device security chip, and message integrity and timeliness are verified. Execution environment isolation is implemented to ensure that operations are completed in a trusted execution environment. Execution status data is collected in real time and the execution results are transmitted back through a secure channel.

[0096] It should be noted that equipment safety levels are divided into critical equipment and ordinary equipment. Critical equipment is equipment that directly affects the safety of the power grid and may cause major accidents, while ordinary equipment is equipment that only affects local power supply and has relatively minor consequences of failure.

[0097] Preferably, through pattern matching and semantic mapping, collaborative scheduling instructions are converted into standardized device control parameters, reducing protocol differences between heterogeneous devices and improving instruction compatibility; combining physical constraint real-time verification, encrypted transmission and identity authentication to enhance execution security; utilizing timing-dependent instruction sorting and multi-mode transmission channel selection to improve and optimize execution efficiency.

[0098] S6. Use dynamic threshold filtering to monitor the equipment operating status, safety boundaries and economic thresholds in real time, and use online learning to optimize equipment control instructions.

[0099] Furthermore, edge computing nodes are used to collect real-time campus equipment data and combined with preprocessing to generate a standardized data matrix.

[0100] Specifically, edge computing nodes are used to collect real-time park equipment data at a fixed frequency and through event triggering (such as sudden changes in equipment status) and temporarily store it in a local cache area. A first-in-first-out strategy is adopted to manage and generate an initial data matrix. Based on the initial data matrix, the window size is automatically selected according to the data fluctuation characteristics. For example, the window length for photovoltaic output data (which fluctuates faster) is 10 seconds, and the window length for load data (which changes slower) is 1 minute. A sliding window calculation is performed independently on each monitoring parameter to obtain filtered smoothed data. Abnormal data is detected and corrected to generate a denoised data matrix. The mean and standard deviation of the data are calculated in real time within the sliding window and standardized. Based on the denoised data matrix, the denoised data is converted into a standard distribution with a mean of 0 and a variance of 1, eliminating the dimensional differences of different parameters to generate a standardized data matrix.

[0101] Through the standardized data matrix, the safety boundary threshold and economic threshold are obtained by real-time working condition analysis, and a dynamic threshold table is constructed.

[0102] Specifically, based on the standardized data matrix, the mean and standard deviation of each parameter (30-day data) in the standardized data matrix are obtained, and the standardized value is obtained using the standardization formula. The safety boundary threshold is set by obtaining the standard range from the equipment manual and adjusting it according to the standardized value. The standardized value remains within the standard range between -1 and 1, and the standard range is tightened by 5% when the standardized value is between -1 and -2 or 1 and 2. When the standardized value is less than -2 or greater than 2, the standard range is tightened by 10%. The economic threshold is set by obtaining time-of-use electricity price information and performing standardization processing, obtaining the 30-day average electricity price and standard deviation, and using the standardized value formula to obtain the standardized electricity price. When the standardized electricity price is greater than 1, electricity consumption is reduced (lower limit +5%), when the standardized electricity price is less than -1, electricity consumption is increased (upper limit -5%), and when the standardized electricity price is greater than -1 and less than 1, the intermediate value is maintained. The safety boundary threshold is combined with the economic threshold, with the safety threshold given priority, and economic optimization is performed within the safety range. For example, if the safety range of the energy storage battery is 20%-90%, the economic recommendation is to charge to 85% during valley charging, and the execution value is 85%. The output is used to construct a dynamic threshold table.

[0103] It should be noted that real-time operating condition analysis achieves dynamic safety thresholds through real-time monitoring and dynamic evaluation of equipment operating data, environmental parameters, grid status and other information, ensuring safe operation, optimizing economic scheduling and improving the adaptability of scheduling strategies.

[0104] It should be noted that based on the dynamic threshold table, federated learning is used to optimize device control instructions, perform quantum verification, and generate a verification report.

[0105] Based on the dynamic threshold table, distributed strategy optimization is used to preliminarily optimize device control instructions.

[0106] Specifically, current operating data is collected from various devices in the park, including the remaining power (percentage) of the energy storage battery, the actual output of photovoltaic power generation, the power value of the grid connection, and the temperature data of each device; historical data in the past 1 hour is extracted to obtain the average value and fluctuation range of the energy storage battery power, the average value and fluctuation range of the photovoltaic power generation output, and a dynamic safety range is set based on the dynamic threshold table. For example, the energy storage battery power fluctuates within the range above and below the average value, and the photovoltaic output does not exceed a certain percentage of the rated power; each sub-site, such as the photovoltaic area and the energy storage area, uses federated reinforcement learning to obtain recommended adjustment values ​​based on the current operating status and the dynamic threshold table, such as the recommended energy storage charge and the recommended photovoltaic output value. Optimization suggestions for each sub-site are collected, and the average of each recommended value is taken as the final optimized value. Continuous values ​​are rounded, such as to 10 kilowatts, and mutually exclusive operations of equipment, such as charging and discharging cannot be performed simultaneously, are processed to generate optimized equipment control instructions.

[0107] According to the optimized equipment control instructions, multimodal simulation verification is used to perform quantum verification and digital twin testing, and a verification report is generated.

[0108] Specifically, the optimization device control instructions are converted into binary codes, and combined with the constraints in the source-grid-load-storage-charging optimization model, quantum computing is used to verify whether the instruction combination violates the constraints, such as checking whether the energy storage charging and discharging instructions exceed the SOC limit, verifying whether the photovoltaic output exceeds the maximum allowable value, and generating a quantum verification report; based on the device parameters in the source-grid-load-storage-charging optimization model, such as the charging and discharging efficiency of the energy storage, the output characteristics of the photovoltaic array, and the grid connection parameters, control instructions are executed in a digital environment, and voltage stability, equipment operating status, and energy supply and demand balance are monitored. Abnormal conditions during operation are identified, the time of occurrence of the abnormality, the type of abnormality, and the scope of impact are recorded, and a digital twin test report is generated; the quantum verification report and the digital twin test report are compared to evaluate the safety and feasibility of the optimization device control instructions, extract the list of verified instructions, instructions that need to be adjusted, and adjustment suggestions, conduct an overall operation risk assessment, and generate a verification report.

[0109] Based on the verification report, a hierarchical execution strategy and feedback optimization are adopted to generate optimized equipment control instructions.

[0110] Specifically, based on the verification report, the actual operating voltage value of the equipment, the actual operating current value of the equipment, the power output value of the equipment, and the grid frequency change value after executing the instruction are extracted, the equipment response time and the statistical instruction execution success rate (percentage) are recorded, and an equipment operation data record table containing measured values ​​and timestamps is generated; the measured values ​​are compared with the rated parameters of the equipment, and if the voltage deviation exceeds ±5%, it is a first-level instruction for immediate adjustment, and if the current deviation exceeds ±10%, it is a second-level instruction for rapid adjustment, and if the power deviation is within ±5%, it is a third-level instruction for maintenance observation, and it is adjusted according to the grid frequency fluctuation. If the frequency change is greater than 0.2Hz, the adjustment level is increased, and if the frequency change is less than 0.1Hz, the adjustment level is reduced, and the instruction hierarchical execution table is output; the fixed step adjustment method is used to obtain the current value and the target value The first-level instruction is adjusted by 5% each time, the second-level instruction is adjusted by 3% each time, and the third-level instruction is adjusted by 1% each time. The restriction conditions can be added as follows: the voltage adjustment does not exceed ±10%, the current adjustment does not exceed ±15%, and the number of adjustments per day does not exceed 10 times, and the equipment parameter adjustment instructions are generated; the actual operation data after each adjustment is recorded, the voltage stabilization time, current fluctuation range and power recovery speed are obtained, the number of equipment failures is counted, the grid frequency fluctuation is recorded, and an adjustment effect record table with before-and-after comparison data is generated; the adjustment range with the best effect is counted, the adjustment operation that causes the failure is identified, the parameter combination with the best grid response is recorded, the fixed step value is optimized, the classification standard is adjusted, and the daily adjustment upper limit is modified to generate optimized equipment control instructions.

[0111] The better ones use dynamic threshold monitoring, quantum verification and digital twin testing to enhance safety; use dynamic economic threshold adjustment, time-of-use electricity prices and load demand to optimize energy storage charging and discharging strategies in real time and optimize economic scheduling; continuously improve scheduling accuracy through federated reinforcement learning dynamic improvement strategies, handle them in a graded manner according to the degree of urgency, balance real-time performance and stability, and improve adaptability.

[0112] This embodiment also provides a computer device, which is suitable for the AI-based intelligent park source-grid-load-storage-charging integrated scheduling method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the AI-based intelligent park source-grid-load-storage-charging integrated scheduling method proposed in the above embodiment.

[0113] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0114] This embodiment also provides a storage medium having a computer program stored thereon. When the program is executed by a processor, it implements the AI-based intelligent park source-grid-load-storage-charging integrated scheduling method proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0115] In summary, the present invention achieves this by: combining a hybrid neural network architecture with quantum state spatial feature extraction to deeply explore dynamic correlation features between devices and enhance the characterization capability of spatiotemporal features; and using proximal strategy to optimize the pruning algorithm and Nash equilibrium game to enhance the economic efficiency of the scheduling strategy.

[0116] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An AI-based intelligent park integrated source-grid-load-storage-charging scheduling method, characterized by: include, Collect and pre-process park equipment data in real time to generate a standardized time series data matrix; Use time series information to filter and extract data features of the standardized time series data matrix, and use a hybrid neural network framework to build a source-grid-load-storage optimization model to obtain an embedding vector; The embedded vector is mapped to the action space through the policy network, and the proximal policy optimization and pruning algorithm is used to obtain the preliminary scheduling strategy; Use game theory to optimize the preliminary scheduling strategy and generate a collaborative scheduling instruction set. Specifically, consider the park equipment as game participants, define the payoff function of each device, output the game parameters, and map the Nash equilibrium solution set to the target space based on the game parameters to generate the Pareto frontier. Use the ε-constraint method to impose constraints on the secondary objectives based on the primary objective, screen non-inferior solutions, output the Pareto optimal solution, screen the optimization adjustment strategy, use the semantic template to map the optimization adjustment strategy to the equipment control parameters, and generate a collaborative scheduling instruction set. Using protocol conversion, the collaborative scheduling instruction set is converted into device control instructions and executed; Dynamic threshold filtering is used to monitor equipment operating status, safety boundaries and economic thresholds in real time, and online learning is used to optimize equipment control instructions.

2. The AI-based intelligent park integrated source-grid-load-storage-charging scheduling method according to claim 1, characterized in that: The real-time collection of park equipment data and preprocessing to generate a standardized time series data matrix are as follows: Time synchronization is performed through the time protocol, and the sampling frequency is adjusted using the dynamic sampling rate to collect data from campus equipment; Based on the park equipment data, sliding window smoothing and denoising are adopted, and dynamic Z score is used for standardization to construct standardized park equipment data; The standardized campus equipment data is cleaned using linear interpolation completion and anomaly detection, and the cleaned campus equipment data is converted into a standardized time series data matrix through time series alignment.

3. The AI-based intelligent park integrated source-grid-load-storage-charging scheduling method according to claim 2, characterized in that: The method uses time series information to filter and extract the data features of the standardized time series data matrix, and uses a hybrid neural network framework to build a source-grid-load-storage optimization model to obtain an embedded vector. The specific steps are as follows: A hybrid neural network architecture is used to combine spatiotemporal encoders, feature extractors, and decision makers to build a source-grid-load-storage optimization model. The spatiotemporal encoder uses a bidirectional long short-term memory network to perform spatiotemporal position encoding, converting the standardized time series data matrix into an enhanced time series feature matrix; The feature extractor is based on the enhanced temporal feature matrix, combined with Hilbert space projection and causal entanglement gating, to extract quantum attention features in the quantum state space; Based on the quantum attention features, the device association matrix is ​​generated by quantum probability statistics method, and the embedding vector is obtained in combination with the decision maker.

4. The AI-based intelligent park source-grid-load-storage-charging integrated scheduling method according to claim 3 is characterized by: The embedded vector is mapped to the action space through the policy network, and the proximal policy optimization and pruning algorithm is used to obtain the preliminary scheduling strategy. The specific steps are as follows: A three-branch fully connected neural network is used to map the embedding vector to the initial action space and generate the initial action vector; Use masked temperature Softmax combined with device physical constraints to obtain the feasible strategy probability of the initial action vector; Based on the feasible strategy probability, the proximal strategy optimization pruning method and L2 projection correction are combined to optimize the feasible strategy probability, and the Lagrange multiplier method is used to generate the preliminary scheduling strategy.

5. The AI-based intelligent park source-grid-load-storage-charging integrated scheduling method according to claim 4 is characterized by: The game theory is used to optimize the preliminary scheduling strategy and generate a collaborative scheduling instruction set. The specific steps are as follows: Use game theory to adjust and optimize the initial scheduling strategy, and use multi-objective optimization to screen the non-inferior solution set and generate the optimized adjustment strategy; Using semantic templates, the optimization adjustment strategy is mapped to device control parameters and a collaborative scheduling instruction set is generated.

6. The AI-based intelligent park integrated source-grid-load-storage-charging scheduling method according to claim 5, characterized in that: The protocol conversion is used to convert the collaborative scheduling instruction set into device control instructions and execute them. The specific steps are as follows: Use structured data parsing to parse the collaborative scheduling instruction set, extract the device number, control parameters and execution time, and generate device control instructions; Use industrial protocol converter to convert device control instructions into communication protocol format; Use data encryption transmission to send and execute device control instructions through a secure channel.

7. The AI-based intelligent park integrated source-grid-load-storage-charging scheduling method according to claim 6, characterized in that: The dynamic threshold filtering is used to monitor the equipment operation status, safety boundary and economic threshold in real time, and online learning is used to optimize the equipment control instructions. The specific steps are as follows: Use edge computing nodes to collect real-time campus equipment data and combine it with preprocessing to generate a standardized data matrix; Through standardized data matrix, real-time working condition analysis is used to obtain safety boundary thresholds and economic thresholds, and a dynamic threshold table is constructed; Based on the dynamic threshold table, federated learning is used to optimize device control instructions, perform quantum verification, and generate a verification report. Adopt hierarchical execution strategy and feedback optimization to generate optimized equipment control instructions.

8. The AI-based intelligent park integrated source-grid-load-storage-charging scheduling method according to claim 7, characterized in that: The specific steps of using federated learning to optimize device control instructions, perform quantum verification, and generate a verification report are as follows: Based on the dynamic threshold table, distributed strategy optimization is used to preliminarily optimize device control instructions; According to the optimized equipment control instructions, multimodal simulation verification is used to perform quantum verification and digital twin testing, and a verification report is generated.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the AI-based smart park source-grid-load-storage-charging integrated scheduling method described in any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the AI-based smart park source-grid-load-storage-charging integrated scheduling method according to any one of claims 1 to 8 are implemented.

Images