A photovoltaic power generation power prediction and distribution area energy storage microgrid scheduling method and system

Abstract

This invention discloses a method and system for photovoltaic power generation prediction and distribution area energy storage microgrid scheduling, relating to the technical fields of photovoltaic power generation prediction and distribution area energy storage. The method includes: acquiring multi-source operation data of the distribution area microgrid and extracting feature engineering data; inputting the features into an adaptive hybrid model containing deep neural networks and LightGBM to obtain prediction results, and using an alternating optimization strategy and sequential Bayesian updates to achieve online updates; performing multi-timescale optimized scheduling based on the prediction results, generating a day-ahead baseline plan, intraday rolling instructions with added voltage safety penalties, and real-time smoothing control instructions; acquiring actual operation performance data after executing the instructions; and triggering online incremental training of the prediction framework when the prediction error exceeds preset conditions, and dynamically adjusting the safety sub-factor coefficients in the scheduling. This invention breaks down the information barriers between prediction and scheduling, forming an adaptive closed-loop control, significantly improving the economy and safety of microgrid operation.

Description

Technical Field

[0001] This invention relates to the technical field of photovoltaic power generation prediction and distribution area energy storage, and more specifically, to a method and system for photovoltaic power generation prediction and distribution area energy storage microgrid scheduling. Background Technology

[0002] With the rapid increase in the penetration rate of renewable energy sources, represented by photovoltaics, in distribution networks, especially at the substation (power supply range of distribution transformers) level, the significant intermittency, randomness, and volatility of photovoltaic power generation have brought a series of severe challenges to the safe and stable operation of the substation power grid. To solve these problems, configuring substation energy storage systems has become the mainstream solution. However, existing photovoltaic forecasting and energy storage dispatching technologies have obvious limitations in practical applications.

[0003] On the one hand, current photovoltaic power forecasting is mostly concentrated at the regional or power plant level, and the forecasting models often rely on single numerical weather predictions and fixed historical data. Due to the complex and variable local meteorological microenvironment in the distribution area and the lack of accurate monitoring data, traditional forecasting methods suffer from low accuracy and poor timeliness at the distribution area level, failing to provide a reliable basis for real-time dispatch. On the other hand, most current energy storage in distribution areas adopts simple "peak shaving and valley filling" or charging and discharging strategies based on fixed thresholds. This passive response dispatching fails to achieve deep synergy with photovoltaic output forecasting, and cannot fully utilize the regulation potential of energy storage to achieve optimal economic operation of the distribution area.

[0004] More critically, in existing technologies, photovoltaic forecasting modules and energy storage scheduling modules are typically independent or only superficially coupled. Forecast results are transmitted to the scheduling system in the form of a simple power curve, which the system then uses for optimization calculations. This unidirectional, pipeline-like architecture suffers from fatal flaws: one-way information flow and delayed decision-making. When forecasts deviate, the scheduling system cannot correct them in a timely manner, nor can it use the actual deviations in scheduling execution to guide the optimization of the forecasting model. This fragmented system architecture can lead to erroneous charging and discharging commands, not only failing to improve the renewable energy absorption rate but also exacerbating operational risks such as power backfeeding, voltage exceeding limits, and transformer overload in the distribution area. Summary of the Invention

[0005] The purpose of this invention is to provide an integrated solution that can deeply integrate photovoltaic power generation prediction and distribution area energy storage microgrid scheduling, so as to solve the core technical problems of disconnect between prediction and scheduling and lack of adaptive evolution capability of the model in the existing technology.

[0006] The technical solution of this invention is: a method for photovoltaic power generation prediction and distribution area energy storage microgrid scheduling, which is executed by an intelligent main control unit deployed on the distribution area side, and includes:

[0007] S1. Acquire real-time microgrid operation data from the data acquisition layer. The operation data includes photovoltaic power generation collected by the photovoltaic inverter, energy storage system state of charge and charging / discharging power collected by the energy storage converter, load power and node voltage and current collected by the smart meter, and temperature and irradiance collected by the micro weather station; clean and align the operation data, and extract feature engineering data.

[0008] S2. Construct a photovoltaic power prediction framework including a Long Short-Term Memory (LSTM) network and a LightGBM model. During the online model update phase, new data is received through a sliding window mechanism. An alternating optimization strategy is adopted to calculate the gradient of the joint objective function consisting of the prediction error term and the regularization penalty term with respect to the LSTM network parameters under the condition of fixed LightGBM model parameters, so as to update the LSTM network parameters. Sequential Bayes is used to update and smooth the parameter distribution of the LSTM network. Based on the updated LSTM network parameters, fused features are generated. Under the condition of fixed LSTM network parameters, a new decision tree is trained based on the fused features and pseudo-residuals with a preset forgetting factor, using gradient basis one-sided sampling and histogram optimization strategies to update the LightGBM model parameters and obtain the current optimal model. In the prediction phase, feature engineering data is input into the current optimal model to obtain the photovoltaic power prediction results.

[0009] S3. Based on photovoltaic power prediction results, load prediction data, and the state of charge of the energy storage system, perform multi-time-scale optimized scheduling, generate energy storage charging and discharging commands, and issue them to the energy storage converter in the execution control layer. Multi-time-scale optimized scheduling includes: generating a day-ahead baseline charging and discharging plan with the goal of minimizing the total daily operating cost of the distribution area; adding a voltage safety penalty with a safety sub-factor to the day-ahead baseline charging and discharging plan to generate intraday rolling charging and discharging commands; and performing real-time smoothing and voltage emergency support control operations based on the instantaneous values ​​of node voltage and net power.

[0010] S4: Obtain the actual operating effect data of the energy storage converter after executing the energy storage charging and discharging command, and calculate the prediction error and scheduling cost deviation. When the prediction error exceeds the preset conditions, use the actual operating effect data as a feedback signal to perform online incremental training on the photovoltaic power generation prediction framework, and dynamically adjust the safety sub-coefficient in the multi-time scale optimization scheduling to form a closed-loop control of prediction and scheduling.

[0011] In any of the above technical solutions, further, in step S2, the specific process of updating the parameters of the Long Short-Term Memory (LSTM) network and the LightGBM model includes:

[0012] Receive the Approximate new data Update the sliding window to ,in The length of the sliding window;

[0013] With fixed LightGBM parameters Under these conditions, the parameters of the Long Short-Term Memory (LSTM) network are updated using gradient descent. : ;in, The learning rate for the Long Short-Term Memory (LSTM) network. The joint objective function consists of a prediction error term and a regularization penalty term.

[0014] Smooth parameter distribution by sequential Bayes:

[0015] ;

[0016] Calculate fusion features based on updated Long Short-Term Memory (LSTM) network parameters. ,in Indicates For input variables, with The output features of the Long Short-Term Memory (LSTM) network for model parameters. To concatenate the weight matrix;

[0017] With fixed parameters of the Long Short-Term Memory (LSTM) network Under the given conditions, calculate the forgetting factor. pseudo residuals :

[0018] ;

[0019] Based on fusion features and pseudo residuals Training a new decision tree Optimize the weight of new trees And update the LightGBM model to .

[0020] In any of the above technical solutions, step S2 further includes an online adaptive prediction stage:

[0021] At the beginning of each scheduling cycle, real-time data and short-term numerical weather forecasts are input into the current optimal time series model to obtain the future photovoltaic power time series prediction sequence value.

[0022] At a specific time point on the day, a parallel non-time series model is invoked to obtain the non-time series prediction sequence for the day, and the error is compared between the non-time series prediction sequence and the time series prediction sequence values.

[0023] Using the prediction error from the previous moment as input, a first-order autoregressive error correction model is used to dynamically weight and correct the current prediction value, and the final photovoltaic power prediction result is output.

[0024] In any of the above technical solutions, further, in step S3, the multi-timescale optimization scheduling specifically includes: a day-ahead optimization scheduling layer, an intraday rolling optimization layer, and a real-time control layer;

[0025] The current optimization scheduling layer is as follows: Input the next day's weather forecast, photovoltaic forecast curve, load forecast curve, time-of-use electricity price curve, and energy storage system parameters; establish a mixed-integer programming model with the objective of minimizing the total daily operating cost of the distribution area, with the objective function being:

[0026] ;

[0027] in, For time-of-use electricity pricing for the power grid, For grid interaction power, Costs associated with energy storage charging and discharging losses. For energy storage charging and discharging power;

[0028] Intraday Rolling Optimization Layer: Based on the day-ahead baseline charge / discharge plan, a rolling optimization model is established, with the objective function being:

[0029] ;

[0030] in, Given the economic cost data, This is the voltage over-limit penalty coefficient. This represents the instantaneous value of the node voltage. This is the rated reference voltage;

[0031] Real-time control layer: adopts proportional-integral or fuzzy control rules, and smooths grid-connected power through a first-order low-pass filter; when a node voltage instantaneously exceeds the limit, it calculates the required reactive or active power support from energy storage based on sensitivity analysis and executes it first.

[0032] A scheduling system applying the photovoltaic power generation prediction and distribution area energy storage microgrid scheduling method described in any of the above technical solutions is also provided. This system includes:

[0033] The data acquisition layer is used to collect real-time microgrid operation data in the distribution area. The data acquisition layer includes: photovoltaic inverters for collecting photovoltaic power generation; energy storage converters for collecting the state of charge and charging / discharging power of the energy storage system; smart meters for collecting load power and node voltage and current; and micro weather stations for collecting temperature and irradiance.

[0034] The communication network layer is used to upload the operational data acquired by the data acquisition layer in real time and issue scheduling instructions.

[0035] The intelligent main control unit, connected to the communication network layer, includes: a data processing module for cleaning, aligning, and extracting features from operational data; a prediction module for obtaining photovoltaic power prediction results based on an adaptive hybrid model composed of a Long Short-Term Memory (LSTM) network and a LightGBM model; an optimization scheduling module for performing multi-timescale optimization scheduling based on photovoltaic power prediction results and generating energy storage charging and discharging commands; and a closed-loop feedback module for acquiring actual operational performance data, performing online incremental training of the prediction module when the prediction error exceeds preset conditions, and dynamically adjusting the safety factor of the optimization scheduling module.

[0036] The execution control layer, including the energy storage converter, is used to receive energy storage charging and discharging commands issued by the intelligent main control unit and control the energy storage battery to perform charging and discharging actions.

[0037] The beneficial effects of this invention are:

[0038] Compared with existing technologies, this invention provides a photovoltaic power generation prediction and distribution area energy storage microgrid scheduling method and system. By constructing a closed-loop intelligent master control architecture of prediction-scheduling-evaluation-feedback-re-prediction, it breaks down the information barrier between prediction and scheduling in the traditional mode. After obtaining the actual operating effect data after the energy storage converter executes the command, this invention calculates the deviation between prediction error and scheduling cost. When the prediction error exceeds a preset condition, the actual operating effect data is used as a feedback signal to trigger online incremental training of the photovoltaic power generation prediction framework, while dynamically adjusting the safety component coefficients in the multi-timescale optimization scheduling. This closed-loop feedback mechanism enables the prediction model to evolve online based on the actual scheduling effect. The scheduling strategy also becomes more accurate and proactive due to the more reliable prediction information and dynamic safety constraints, realizing a shift from passive response to proactive guidance, and significantly improving the overall system's operating efficiency and stability.

[0039] In the photovoltaic power generation prediction stage, this invention employs an adaptive hybrid model composed of a deep neural network and a LightGBM model, effectively integrating local microclimate data and real-time operational data. During the online model update process, this invention introduces a sliding window mechanism and an alternating optimization strategy. Under the condition of fixing the parameters of one model, gradients are calculated to update the parameters of the other model, and sequential Bayesian updates are used to smooth the parameter distribution of the deep neural network. This technique effectively avoids parameter mutations and model collapse problems during joint training of complex hybrid models, fully leveraging the high-dimensional feature extraction capability of deep neural networks for complex nonlinear correlations and the efficient interpretability of tree models. This significantly improves the model's adaptability to the complex and variable microclimate environment of the power distribution area, and greatly enhances the prediction accuracy of ultra-short-term photovoltaic power.

[0040] This invention implements a multi-timescale optimized scheduling strategy during the energy storage dispatch phase, achieving a dynamic balance between the economic efficiency and safety of distribution network operation. At the day-ahead optimization dispatch layer, a baseline charge-discharge plan is generated with the goal of minimizing the total daily operating cost of the distribution network, maximizing peak-valley arbitrage benefits and improving the local photovoltaic (PV) consumption rate. At the intraday rolling optimization layer, based on ultra-short-term forecast results, a voltage safety penalty term including a safety factor is added to the day-ahead plan, proactively preventing voltage overruns and transformer overloads caused by PV backfeed power. At the real-time control layer, based on the instantaneous values ​​of node voltage and net power, a first-order low-pass filter smooths the grid-connected power, and emergency support control is prioritized when voltage exceeds limits. This hierarchical dispatch mechanism enables the distribution network microgrid to respond quickly to system disturbances, effectively resist load surges and PV fluctuations, and reduce users' electricity costs while ensuring local power supply quality. Attached Figure Description

[0041] The advantages of the above and additional aspects of the present invention will become apparent and readily understood in the description of the embodiments in conjunction with the following drawings, wherein:

[0042] Figure 1 This is a schematic flowchart of a photovoltaic power generation prediction and distribution area energy storage microgrid scheduling method and system according to an embodiment of the present invention;

[0043] Figure 2 This is a schematic diagram of the system structure of a photovoltaic power generation prediction and distribution area energy storage microgrid scheduling method and system according to an embodiment of the present invention. Detailed Implementation

[0044] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0045] In the following description, many specific details are set forth in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0046] like Figure 2 As shown, this embodiment provides a photovoltaic power generation prediction and distribution area energy storage microgrid dispatching system. The system uses an integrated intelligent main control unit as its core, deployed in the distribution room or cloud server on the distribution area side. The system includes: a data acquisition layer, a communication network layer, an intelligent main control unit, and an execution control layer. The specific composition and parameter interaction relationships of each layer are as follows:

[0047] Data Acquisition Layer: Used for real-time acquisition of multi-dimensional physical parameters of the microgrid operation in the distribution area. This layer specifically includes: photovoltaic inverters, used to acquire real-time photovoltaic power generation. Energy storage converter, used to collect the state of charge (SOC) and charging / discharging power of the energy storage system. Smart meters are used to collect load power. Voltage at each node and current Miniature weather stations are used to collect real-time and future temperature and irradiance data; intelligent fusion terminals in the distribution area are used to aggregate the status of the aforementioned underlying equipment.

[0048] Communication network layer: Using wired communication methods such as optical fiber or power line carrier, or wireless communication methods such as 4G, 5G, or LoRa, it is responsible for uploading the physical parameters obtained by the data acquisition layer to the intelligent main control unit in real time, and issuing the scheduling instructions generated by the intelligent main control unit.

[0049] Intelligent main control unit: This is the core decision-making center of the system, which includes a data processing module, a prediction module, an optimization scheduling module, and a closed-loop feedback module.

[0050] The execution control layer includes an energy storage converter, which receives charging and discharging power commands from the intelligent main control unit, precisely controls the physical charging and discharging actions of the energy storage battery, and monitors its own operating status in real time.

[0051] Based on the above system architecture, this embodiment of the invention also provides a method for photovoltaic power generation prediction and distribution area energy storage microgrid scheduling. This method is executed by the aforementioned intelligent main control unit, such as... Figure 1 As shown, the specific steps include:

[0052] S1. Multi-source Data Fusion and Processing: The data processing module of the intelligent main control unit receives the underlying data uploaded from the communication network layer and performs cleaning, alignment, normalization, and feature engineering. During the data cleaning phase, missing values ​​caused by communication interruptions are handled (filled by interpolation or predicted values ​​from previous and subsequent time points), and outliers judged based on the 3σ principle or physical thresholds are removed. During the data alignment phase, minute-level meteorological data collected by the micro-weather station and second-level power data collected by the photovoltaic inverter are aggregated into a 1-hour granularity dataset, or unified to a 15-minute scheduling time granularity through interpolation and aggregation methods.

[0053] During the feature engineering phase, the data processing module calculates the historical irradiance moving average, rate of change, and interaction term between temperature and irradiance based on the irradiance data from the micro-weather station; and calculates the photovoltaic power generation based on the photovoltaic power collected by the photovoltaic inverter. Load power collected by smart meters Calculate the net load of the transformer area. It identifies peak and off-peak periods of load and calculates the difference between real-time electricity price and average electricity price.

[0054] S2. Photovoltaic Power Generation Prediction Based on Adaptive Hybrid Model: The prediction module of the intelligent main control unit adopts a hybrid prediction framework of "LightGBM + Long Short-Term Memory Network (LSTM) + Online Learning". The LSTM network serves as a feature extractor to capture complex nonlinear correlations. The specific prediction process includes:

[0055] S21. Initialization Phase: Initialize the parameters of the Long Short-Term Memory (LSTM) network using the Xavier initialization method. Collect at least one year of historical data from the transformer substation as the training set. Based on the dataset output by the data processing module, construct a LightGBM prediction model (including time-series and non-time-series models) with gradient boosting decision trees at its core. Input features include: predicted future irradiance and temperature provided by the micro-weather station, and historical and current photovoltaic output from the photovoltaic inverter, combined with date type and seasonal information. Use initial batch data. The initial LightGBM model is trained with the goal of minimizing the root mean square error, resulting in the initial baseline prediction time series model: To obtain the initial parameters Simultaneously set the sliding window length. Preset forgetting factor Online iteration cycle Learning rate of Long Short-Term Memory (LSTM) network Learning rate with LightGBM .

[0056] in, For constant prediction values, For the first The actual photovoltaic power label of a sample This is the loss function.

[0057] S22, Model Training and Online Iterative Optimization Phase: Receive the first... Approximate new data The sliding window is updated to .

[0058] in, For the first The input feature matrix of the batch data, For the corresponding actual photovoltaic power label, Indicates all belonging to But not belonging to A collection of elements.

[0059] Optimize the parameters of the Long Short-Term Memory (LSTM) network. (Using an alternating optimization strategy, while keeping the LightGBM parameters fixed) Optimize the parameters of the Long Short-Term Memory (LSTM) network while keeping them unchanged. ): Calculate the joint objective function based on the weighted loss of data within the window. The gradient is used to update the parameters via gradient descent. : .

[0060] Wherein, the joint objective function It consists of a prediction error term and a regularization penalty term, and is used to prevent the model from overfitting while ensuring prediction accuracy.

[0061] Simultaneously, the parameter distribution is updated and smoothed using sequential Bayesian methods:

[0062] ;

[0063] Generating fusion features: Based on the updated parameters of the Long Short-Term Memory (LSTM) network, calculate the fusion features of the data within the window:

[0064] ;

[0065] in, Indicated by For input variables, with The output features of the Long Short-Term Memory (LSTM) network for model parameters are shown. The semicolon is used to distinguish between input variables and model parameters. This indicates that the original features and deep features are concatenated; It is a learnable concatenation weight matrix used for linear mapping and dimensionality reduction fusion of the concatenated high-dimensional features.

[0066] Optimize LightGBM parameters (Using an alternating optimization strategy, with fixed parameters of the Long Short-Term Memory (LSTM) network) Optimize LightGBM parameters while keeping them unchanged. ): Calculate the pseudo-residuals with a pre-defined forgetting factor:

[0067] ;

[0068] based on and A new decision tree is trained using gradient basis one-sided sampling and histogram optimization strategies. Optimize the weight of new trees :

[0069] ;

[0070] in, For the current sliding window The set of sample indexes within.

[0071] Update the LightGBM model: .

[0072] S23. Convergence Judgment Phase: When the decrease in the joint objective function for two consecutive rounds... ( (For the convergence threshold), or when the online iteration cycle reaches [a certain number]... When the time is right, stop online updates and output the current optimal model.

[0073] S24. Online Adaptive Prediction Phase: In the prediction phase, at the beginning of each scheduling cycle, real-time data and short-term numerical weather forecasts are input into the current optimal time-series model to obtain the future photovoltaic power time-series prediction sequence. At a specific time point on the day, a parallel non-time-series model is invoked to obtain the non-time-series prediction sequence for that day. By comparing the error with the time-series prediction sequence value, a dynamic weighted correction module is invoked for optimization. Specifically, a standard discrete Kalman filter or a first-order autoregressive error correction model is used, with the prediction error from the previous moment as input to dynamically correct the current prediction value. For example, an error correction formula is used. ,in This is the current predicted value. This represents the actual error at the previous moment. To correct the gain coefficient, the final photovoltaic power prediction result is output.

[0074] S3. SOC Prediction Configuration and Multi-Time-Scale Optimization Scheduling of Energy Storage in Distribution Areas: This step is executed by the optimization scheduling module of the intelligent master control unit, and is specifically divided into two stages: SOC prediction configuration and hierarchical scheduling.

[0075] First, the SOC prediction and configuration for the energy storage area is performed: At the beginning of each scheduling cycle, real-time data and the numerical weather forecast for the next 24 hours are input into the benchmark LightGBM model to obtain the hourly SOC setting strategy sequence values ​​for the next 24 hours. Every hour of the day, the current SOC value of the energy storage converter is obtained and the SOC plan is updated, and verification and matching are performed based on the deviation between the actual SOC and the target. When it is determined that the target SOC cannot be achieved at critical time points, an emergency adjustment with a larger rate of change is used during off-peak hours.

[0076] Subsequently, a hierarchical decision-making framework is adopted to perform multi-time-scale optimization scheduling: the optimization scheduling module of the intelligent master control unit adopts a hierarchical decision-making framework, including a day-ahead optimization scheduling layer, an intraday rolling optimization layer, and a real-time control layer.

[0077] The current-day optimization scheduling layer is executed at a fixed time each day. Inputs include the next day's weather forecast, photovoltaic / load forecast curves, time-of-use electricity price curves, and energy storage system parameters. A mixed-integer programming model is established with the objective of minimizing the total daily operating cost of the distribution area.

[0078] ;

[0079] in, This represents the total number of scheduling cycles within a day. For the current time step, To schedule the time step, For time-of-use electricity pricing for the power grid, Costs associated with energy storage charging and discharging losses. Constraints include power balance constraints. The system includes the dynamic equations and upper and lower limits of the energy storage SOC, as well as transformer capacity constraints. It outputs the energy storage baseline charge and discharge plan for the following day.

[0080] Intraday Rolling Optimization Layer: Executed every 15 minutes or 1 hour. Inputs include the latest ultra-short-term forecast curve, real-time electricity price, actual SOC of the energy storage converter, day-ahead plan, and real-time voltage / current of smart meters. A rolling optimization model is established based on the day-ahead plan, with the objective function adding a voltage safety penalty term including a safety factor to the known day-ahead economic cost data.

[0081] ;

[0082] in, Given the economic cost data, This is the safety factor (also known as the voltage over-limit penalty factor). This is the rated reference voltage. Output updated energy storage charge / discharge commands.

[0083] Real-time control layer: Executes every second or every few seconds. Input data processing module calculates the net power of the transformer area. Instantaneous value, instantaneous voltage value of smart meter, intraday charge / discharge command Fast closed-loop control is achieved using proportional-integral or fuzzy control rules. A first-order low-pass filter is designed to smooth grid-connected power; if a momentary voltage exceedance at a node is detected, the required reactive or active power support from energy storage is calculated based on sensitivity analysis and executed preferentially.

[0084] S4. Closed-Loop Feedback and Model Adaptive Update: The closed-loop feedback module of the intelligent main control unit tracks the execution effect of the control layer's instructions. After each scheduling cycle, it calculates the absolute and relative errors of photovoltaic power prediction; assesses the deviation between actual operating costs and optimized calculation costs; and analyzes the deviation between voltage qualification rate and energy storage SOC trajectory. Feedback reports are generated periodically. When the prediction error remains large, the baseline prediction model is incrementally trained using recent actual data. Simultaneously, based on the voltage control effect feedback from smart meters, the safety sub-factor coefficients in the intraday rolling optimization layer are dynamically adjusted. This forms a closed-loop control system for prediction and scheduling.

[0085] In summary, this invention proposes a method for photovoltaic power generation prediction and distribution area energy storage microgrid scheduling, comprising:

[0086] S1. Acquire real-time microgrid operation data from the data acquisition layer. The operation data includes photovoltaic power generation collected by the photovoltaic inverter, energy storage system state of charge and charging / discharging power collected by the energy storage converter, load power and node voltage and current collected by the smart meter, and temperature and irradiance collected by the micro weather station. Clean and align the operation data, and extract feature engineering data.

[0087] S2. Construct a photovoltaic power prediction framework including a Long Short-Term Memory (LSTM) network and a LightGBM model. During the online model update phase, new data is received through a sliding window mechanism. An alternating optimization strategy is adopted to calculate the gradient of the joint objective function consisting of the prediction error term and the regularization penalty term with respect to the LSTM network parameters under the condition of fixed LightGBM model parameters, so as to update the LSTM network parameters. Sequential Bayes is used to update and smooth the parameter distribution of the LSTM network. Based on the updated LSTM network parameters, fused features are generated. Under the condition of fixed LSTM network parameters, a new decision tree is trained based on the fused features and pseudo-residuals with a preset forgetting factor, using gradient basis one-sided sampling and histogram optimization strategies to update the LightGBM model parameters and obtain the current optimal model. In the prediction phase, feature engineering data is input into the current optimal model to obtain the photovoltaic power prediction results.

[0088] S3. Based on photovoltaic power prediction results, load prediction data, and the state of charge of the energy storage system, perform multi-time-scale optimized scheduling, generate energy storage charging and discharging commands, and issue them to the energy storage converter in the execution control layer. Multi-time-scale optimized scheduling includes: generating a day-ahead baseline charging and discharging plan with the goal of minimizing the total daily operating cost of the distribution area; adding a voltage safety penalty with a safety sub-factor to the day-ahead baseline charging and discharging plan to generate intraday rolling charging and discharging commands; and performing real-time smoothing and voltage emergency support control operations based on the instantaneous values ​​of node voltage and net power.

[0089] S4: Obtain the actual operating effect data of the energy storage converter after executing the energy storage charging and discharging command, and calculate the prediction error and scheduling cost deviation. When the prediction error exceeds the preset conditions, use the actual operating effect data as a feedback signal to perform online incremental training on the photovoltaic power generation prediction framework, and dynamically adjust the safety sub-coefficient in the multi-time scale optimization scheduling to form a closed-loop control of prediction and scheduling.

[0090] The systems upon which the above methods rely include:

[0091] The data acquisition layer is used to collect real-time microgrid operation data in the distribution area. The data acquisition layer includes: photovoltaic inverters for collecting photovoltaic power generation; energy storage converters for collecting the state of charge and charging / discharging power of the energy storage system; smart meters for collecting load power and node voltage and current; and micro weather stations for collecting temperature and irradiance.

[0092] The communication network layer is used to upload the operational data acquired by the data acquisition layer in real time and issue scheduling instructions.

[0093] The intelligent main control unit, connected to the communication network layer, includes: a data processing module for cleaning, aligning, and extracting features from operational data; a prediction module for obtaining photovoltaic power prediction results based on an adaptive hybrid model composed of a Long Short-Term Memory (LSTM) network and a LightGBM model; an optimization scheduling module for performing multi-timescale optimization scheduling based on photovoltaic power prediction results and generating energy storage charging and discharging commands; and a closed-loop feedback module for acquiring actual operational performance data, performing online incremental training of the prediction module when the prediction error exceeds preset conditions, and dynamically adjusting the safety factor of the optimization scheduling module.

[0094] The execution control layer, including the energy storage converter, is used to receive energy storage charging and discharging commands issued by the intelligent main control unit and control the energy storage battery to perform charging and discharging actions.

[0095] The steps in this invention can be adjusted, combined, or deleted according to actual needs.

[0096] The units in the device of the present invention can be merged, divided, or reduced according to actual needs.

[0097] In this invention, the terms "installation," "connection," "linking," and "fixing" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; "linking" can be a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of these terms in this invention according to the specific circumstances.

[0098] The shapes of the components in the accompanying drawings are schematic and may differ from their actual shapes. The drawings are only used to illustrate the principles of the present invention and are not intended to limit the present invention.

[0099] Although the invention has been disclosed in detail with reference to the accompanying drawings, it should be understood that these descriptions are merely exemplary and not intended to limit the application of the invention. The scope of protection of the invention is defined by the appended claims and may include various modifications, alterations, and equivalents made to the invention without departing from the scope and spirit of the invention.

Claims

1. A method for photovoltaic power generation prediction and distribution area energy storage microgrid scheduling, characterized in that, The method is executed by an intelligent main control unit deployed on the transformer substation side, and the method includes: S1. Acquire real-time microgrid operation data from the data acquisition layer. The operation data includes photovoltaic power generation collected by the photovoltaic inverter, energy storage system state of charge and charging / discharging power collected by the energy storage converter, load power and node voltage and current collected by the smart meter, and temperature and irradiance collected by the micro weather station; clean and align the operation data, and extract feature engineering data. S2. Construct a photovoltaic power prediction framework including a Long Short-Term Memory (LSTM) network and a LightGBM model. During the online model update phase, new data is received through a sliding window mechanism. An alternating optimization strategy is adopted to calculate the gradient of the joint objective function consisting of the prediction error term and the regularization penalty term with respect to the LSTM network parameters under the condition of fixed LightGBM model parameters, so as to update the LSTM network parameters. Sequential Bayes is used to update and smooth the parameter distribution of the LSTM network. Based on the updated LSTM network parameters, fused features are generated. Under the condition of fixed LSTM network parameters, a new decision tree is trained based on the fused features and pseudo-residuals with a preset forgetting factor, using gradient basis one-sided sampling and histogram optimization strategies to update the LightGBM model parameters and obtain the current optimal model. In the prediction phase, feature engineering data is input into the current optimal model to obtain the photovoltaic power prediction results. S3. Based on photovoltaic power prediction results, load prediction data, and the state of charge of the energy storage system, perform multi-time-scale optimized scheduling, generate energy storage charging and discharging commands, and issue them to the energy storage converter in the execution control layer. Multi-time-scale optimized scheduling includes: generating a day-ahead baseline charging and discharging plan with the goal of minimizing the total daily operating cost of the distribution area; adding a voltage safety penalty with a safety sub-factor to the day-ahead baseline charging and discharging plan to generate intraday rolling charging and discharging commands; and performing real-time smoothing and voltage emergency support control operations based on the instantaneous values ​​of node voltage and net power. S4: Obtain the actual operating effect data of the energy storage converter after executing the energy storage charging and discharging command, and calculate the prediction error and scheduling cost deviation. When the prediction error exceeds the preset conditions, use the actual operating effect data as a feedback signal to perform online incremental training on the photovoltaic power generation prediction framework, and dynamically adjust the safety sub-coefficient in the multi-time scale optimization scheduling to form a closed-loop control of prediction and scheduling.

2. The photovoltaic power generation prediction and distribution area energy storage microgrid scheduling method as described in claim 1, characterized in that, In step S2, the specific process of updating the parameters of the Long Short-Term Memory (LSTM) network and the LightGBM model includes: Receive the Approximate new data Update the sliding window to ,in The length of the sliding window; With fixed LightGBM parameters Under these conditions, the parameters of the Long Short-Term Memory (LSTM) network are updated using gradient descent. : ;in, The learning rate for the Long Short-Term Memory (LSTM) network. The joint objective function consists of a prediction error term and a regularization penalty term. Smooth parameter distribution by sequential Bayes: ； Calculate fusion features based on updated Long Short-Term Memory (LSTM) network parameters. ,in Indicates For input variables, with The output features of the Long Short-Term Memory (LSTM) network for model parameters. To concatenate the weight matrix; With fixed parameters of the Long Short-Term Memory (LSTM) network Under the condition of, calculate the forgetting factor pseudo residuals : ； Based on fusion features and pseudo residuals Training a new decision tree Optimize the weight of new trees And update the LightGBM model to .

3. The photovoltaic power generation prediction and distribution area energy storage microgrid scheduling method as described in claim 1, characterized in that, Step S2 also includes an online adaptive prediction phase: At the beginning of each scheduling cycle, real-time data and short-term numerical weather forecasts are input into the current optimal time series model to obtain the future photovoltaic power time series prediction sequence value. At a specific time point on the day, a parallel non-time series model is invoked to obtain the non-time series prediction sequence for the day, and the error is compared between the non-time series prediction sequence and the time series prediction sequence value. Using the prediction error from the previous moment as input, a first-order autoregressive error correction model is used to dynamically weight and correct the current prediction value, and the final photovoltaic power prediction result is output.

4. The photovoltaic power generation prediction and distribution area energy storage microgrid scheduling method as described in claim 1, characterized in that, In step S3, the multi-timescale optimization scheduling specifically includes: a day-ahead optimization scheduling layer, an intraday rolling optimization layer, and a real-time control layer; The current optimization scheduling layer is as follows: Input the next day's weather forecast, photovoltaic forecast curve, load forecast curve, time-of-use electricity price curve, and energy storage system parameters; establish a mixed-integer programming model with the objective of minimizing the total daily operating cost of the distribution area, with the objective function being: ； in, For time-of-use electricity pricing for the power grid, For grid interaction power, Costs associated with energy storage charging and discharging losses. For energy storage charging and discharging power; Intraday Rolling Optimization Layer: Based on the day-ahead baseline charge / discharge plan, a rolling optimization model is established with the objective function as follows: ； in, Given the economic cost data, This is the voltage over-limit penalty coefficient. This represents the instantaneous value of the node voltage. This is the rated reference voltage; Real-time control layer: adopts proportional-integral or fuzzy control rules, and smooths grid-connected power through a first-order low-pass filter; when a node voltage instantaneously exceeds the limit, it calculates the required reactive or active power support from energy storage based on sensitivity analysis and executes it first.

5. A scheduling system applying the photovoltaic power generation prediction and distribution area energy storage microgrid scheduling method according to any one of claims 1 to 4, characterized in that, The system includes: The data acquisition layer is used to collect real-time microgrid operation data in the distribution area. The data acquisition layer includes: photovoltaic inverters for collecting photovoltaic power generation; energy storage converters for collecting the state of charge and charging / discharging power of the energy storage system; smart meters for collecting load power and node voltage and current; and micro weather stations for collecting temperature and irradiance. The communication network layer is used to upload the operational data acquired by the data acquisition layer in real time and issue scheduling instructions. The intelligent main control unit, connected to the communication network layer, includes: a data processing module for cleaning, aligning, and extracting features from operational data; a prediction module for obtaining photovoltaic power prediction results based on an adaptive hybrid model composed of a Long Short-Term Memory (LSTM) network and a LightGBM model; an optimization scheduling module for performing multi-timescale optimization scheduling based on photovoltaic power prediction results and generating energy storage charging and discharging commands; and a closed-loop feedback module for acquiring actual operational performance data, performing online incremental training of the prediction module when the prediction error exceeds preset conditions, and dynamically adjusting the safety factor of the optimization scheduling module. The execution control layer, including the energy storage converter, is used to receive energy storage charging and discharging commands issued by the intelligent main control unit and control the energy storage battery to perform charging and discharging actions.

Images