New energy power distribution strategy decision optimization management method based on deep learning

Abstract

The invention relates to the technical field of power system operation optimization, in particular to a new energy power distribution strategy decision optimization management method based on deep learning, which comprises the following steps: acquiring historical supply and consumption data through a power data acquisition network, and performing multi-dimensional data alignment and time sequence coupling analysis to generate joint data. And analyzing the data to determine an optimal input window and a prediction step length of the supply and consumption prediction model, and further executing prediction in parallel to obtain a future supply and consumption prediction set. And carrying out multi-round strategy simulation and evaluation on the prediction sets by utilizing a strategy decision optimization network, and screening out a strategy with the highest internal evaluation score as a target power distribution strategy. According to the method, the prediction precision is improved by enhancing the data time sequence relevance, and the dynamic adaptability and decision robustness of an allocation strategy are enhanced through a closed-loop simulation optimization mechanism.

Description

Technical Field

[0001] This invention relates to the field of power system operation optimization technology, and in particular to a deep learning-based method for optimizing and managing new energy power allocation strategies. Background Technology

[0002] In the field of power system dispatching with a high proportion of renewable energy sources, existing technologies mainly rely on independent time series forecasts for power allocation decisions. Conventional methods model and forecast historical data from both the energy supply and power consumption sides separately, typically using fixed time windows or empirical step sizes. These technologies treat supply and consumption as separable subsystems, lacking systematic processing of the spatiotemporal consistency between multi-source data in their data preprocessing stage. The selection of input variables and step sizes for the forecasting model often depends on prior knowledge, failing to dynamically extract optimal temporal dependencies from the data itself.

[0003] Existing allocation strategies are mostly generated based on prediction results, using static optimization models such as linear programming and heuristic rules for computation. These models rely on deterministic input scenarios, and their decision-making process is essentially a single, open-loop optimization. They cannot simulate and evaluate the cascading effects and long-term performance of allocation strategies under various uncertain future scenarios during the decision-making stage, and lack a strategy learning mechanism that can dynamically interact with the environment and iteratively improve itself through feedback. This results in insufficient robustness and adaptability of existing solutions when facing the strong volatility of new energy sources and the randomness of load.

[0004] There is a need for a method that can improve the quality and correlation of data used for prediction from the source and enable closed-loop learning and optimization of allocation strategies in a simulation environment to cope with the complex uncertainties in system operation. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a deep learning-based method for optimizing the decision-making and management of new energy power allocation strategies.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a deep learning-based new energy power allocation strategy decision optimization management method, comprising: Through the power data acquisition network, acquire historical power supply data sets of multiple energy nodes and historical power consumption data sets of multiple power consumption units within a preset historical period; Multidimensional data alignment and time-series coupling analysis are performed on the historical power supply data set and the historical power consumption data set to generate joint supply and consumption data; In the deep prediction model construction space, fluctuation pattern clustering and feature correlation analysis are performed on the joint supply and consumption data to determine the optimal input variable window and prediction step size for the energy supply prediction model and the consumption prediction model. Based on the optimal input variable window and prediction step size, the energy supply prediction model and the consumption prediction model are run in parallel to obtain future supply prediction data sets for multiple energy nodes and future consumption prediction data sets for multiple power consumption units, respectively. Using a strategy decision optimization network, multiple rounds of strategy simulation and evaluation are performed on the future supply forecast data set and the future consumption forecast data set. The candidate allocation strategy with the highest evaluation score within the strategy decision optimization network is selected as the target power allocation strategy.

[0007] As a further aspect of the present invention, the multi-dimensional data alignment and time-series coupling analysis of the historical power supply data set and the historical power consumption data set includes: Data cleaning is performed on the historical power supply data of each energy node in the historical power supply data set to generate a cleaned supply data sequence. Data cleaning is performed on the historical power consumption data of each power consumption unit in the historical power consumption dataset to generate a cleaned consumption data sequence. Establish a unified timestamp index, and perform data slicing and alignment on the cleaned supply data sequence and the cleaned consumption data sequence based on the unified timestamp index to generate time-aligned supply and consumption data; The time-aligned supply and consumption data are subjected to time-series correlation calculation and mutual information analysis to generate the joint supply and consumption data.

[0008] As a further aspect of the present invention, the step of performing fluctuation pattern clustering and feature correlation analysis on the joint supply and consumption data in the deep prediction model construction space includes: Extract the fluctuation feature vectors of the supply data portion and the consumption data portion from the joint supply and consumption data; The fluctuation feature vectors of the supply data portion and the fluctuation feature vectors of the consumption data portion are respectively input into independent clustering analysis engines; The clustering analysis engine performs clustering analysis on the fluctuation feature vectors of the supply data portion and the consumption data portion respectively, and identifies multiple fluctuation patterns. Calculate the conversion probability between different fluctuation patterns, and construct supply forecast model configuration tables and consumption forecast model configuration tables by combining historical forecast error data.

[0009] As a further aspect of the present invention, determining the optimal input variable window and prediction step size for the energy supply prediction model and the consumption prediction model includes: By iterating through the various fluctuation patterns in the supply forecasting model configuration table, with the goal of minimizing the forecasting error, the optimal input variable window and forecasting step size of the energy supply forecasting model are searched. The system iterates through various fluctuation patterns in the consumption prediction model configuration table, searching for the optimal input variable window and prediction step size of the consumption prediction model with the goal of minimizing the prediction error. The optimal input variable window and prediction step size of the energy supply prediction model and the optimal input variable window and prediction step size of the consumption prediction model obtained from the search are stored in the supply prediction model configuration table and the consumption prediction model configuration table, respectively.

[0010] As a further aspect of the present invention, the parallel operation of the energy supply prediction model and the consumption prediction model includes: According to the supply forecasting model configuration table, configure the current input variable window and forecasting step size of the energy supply forecasting model, and input the supply data of the energy nodes acquired in real time into the energy supply forecasting model. According to the consumption prediction model configuration table, configure the current input variable window and prediction step size of the consumption prediction model, and input the consumption data of the real-time power consumption unit into the consumption prediction model. The energy supply forecasting model outputs the set of future supply forecast data based on its input supply data. The consumption prediction model outputs the set of future consumption prediction data based on its input consumption data.

[0011] As a further aspect of the present invention, the step of utilizing a strategy decision optimization network to perform multiple rounds of strategy simulation and strategy evaluation on the future supply forecast data set and the future consumption forecast data set includes: Initialize the policy decision optimization network, which includes a policy generation module, a power network simulation module, and a policy evaluation module; The strategy generation module generates multiple initial power allocation strategies based on the future supply forecast data set and the future consumption forecast data set. For each initial power allocation strategy, the power network simulation module simulates its execution process within a preset future time period and outputs the simulated supply and consumption trajectory. The strategy evaluation module analyzes the simulated supply and consumption trajectory based on preset evaluation indicators and calculates the evaluation score for each initial power allocation strategy.

[0012] As a further aspect of the present invention, the step of selecting the candidate allocation strategy with the highest evaluation score within the strategy decision optimization network includes: The strategy decision optimization network sorts the evaluation scores of all initial power allocation strategies and selects one or more initial power allocation strategies with the highest evaluation scores as the elite strategy set. The strategy generation module generates new power allocation strategies based on the elite strategy set through strategy crossover and strategy mutation operations, thus forming a new generation of strategy set; For each new power allocation strategy in the next-generation strategy set, the strategy decision optimization network repeatedly executes the power network simulation and the strategy evaluation process to calculate a new evaluation score. The process of strategy evaluation, elite strategy selection, strategy generation and simulation is executed iteratively until the preset iteration termination condition is reached. The strategy with the highest evaluation score in the elite strategy set after the last iteration is determined as the candidate allocation strategy.

[0013] As a further aspect of the present invention, the method further includes model refinement and parameter calibration of the policy decision optimization network: The target power allocation strategy is executed in the actual power network, and the actual supply data and actual consumption data generated during the actual execution process are collected. Calculate the supply forecast deviation between the actual supply data and the future supply forecast data set, and the consumption forecast deviation between the actual consumption data and the future consumption forecast data set; Based on the supply forecast deviation and the consumption forecast deviation, the internal parameters of the energy supply forecast model and the consumption forecast model are adjusted in reverse. The actual execution effect of the target power allocation strategy is compared with the evaluation score given by the strategy decision optimization network in the simulation, and the evaluation index weights of the strategy evaluation module are adjusted according to the comparison results.

[0014] As a further aspect of the present invention, the data cleaning includes: Identify missing and abnormal data points in the historical power supply data set and the historical power consumption data set; For missing data points, data from time points adjacent to the missing data points are used to fill them using time series interpolation methods; For abnormal data points, a statistical distribution-based method is used to replace data points that exceed a preset reasonable range with statistical data from the historical power supply data set or the historical power consumption data set in the corresponding historical period.

[0015] As a further aspect of the present invention, the construction of the supply forecasting model configuration table and the consumption forecasting model configuration table also includes: Associate the supply forecasting model configuration table and the consumption forecasting model configuration table to establish a supply-consumption pattern mapping relationship; The supply-consumption pattern mapping relationship is used to record the joint probability that the consumption data part will exhibit a specific fluctuation pattern when the supply data part is in a specific fluctuation pattern. The supply and consumption pattern mapping relationship is used as prior knowledge and input into the strategy decision optimization network to guide the strategy generation module in generating an initial power allocation strategy that is more consistent with historical joint fluctuation patterns.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: By employing multidimensional data alignment and temporal coupling analysis techniques, this method synchronizes the timeline and mines deep relationships within massive amounts of historical electricity data from various sources. This technique eliminates temporal mismatches and scale differences between heterogeneous data sources, accurately constructing a joint representation of supply and consumption over time. This highly consistent and correlated data foundation directly improves the quality and information content of the input data for subsequent prediction models, enabling them to capture more realistic and refined supply and demand interaction patterns, thereby enhancing the accuracy and reliability of supply and consumption forecasts.

[0017] A policy decision optimization network based on deep reinforcement learning is employed, and a multi-round policy simulation and evaluation mechanism is implemented. This network continuously interacts and learns through trial and error in a virtual environment constructed from high-precision prediction data. Candidate policies are generated through the policy network, and their long-term effectiveness is quantitatively scored by the value network. This process achieves a fundamental shift in allocation policy from static rule formulation to dynamic adaptive optimization, enabling the final selected power allocation policy to proactively adapt to various possible future scenarios, effectively balancing multiple objectives such as system economy, stability, and renewable energy absorption rate. This achieves intelligent decision-making and optimized policy output. Attached Figure Description

[0018] Figure 1 This is a flowchart of the deep learning-based new energy power allocation strategy decision optimization management method described in this invention; Figure 2 This is a flowchart of fluctuation pattern clustering and feature correlation analysis; Figure 3 Heatmap of WAE error for supply forecasting model under high volatility and high correlation mode; Figure 4 The trend curve of new energy supply and power load forecast for the next 4 hours; Figure 5Comparison of the adjustment range of LSTM model parameters. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0021] See Figure 1 The system acquires historical power supply data sets from multiple energy nodes and historical power consumption data sets from multiple power consumption units within a preset historical period through a power data acquisition network. Multidimensional data alignment and temporal coupling analysis are performed on these historical power supply and consumption data sets to generate joint supply and consumption data. Fluctuation pattern clustering and feature correlation analysis are then performed on the joint supply and consumption data within a deep prediction model construction space to determine the optimal input variable window and prediction step size for the energy supply and consumption prediction models. Based on the optimal input variable window and prediction step size, the energy supply and consumption prediction models are run in parallel to obtain future supply prediction data sets from multiple energy nodes and future consumption prediction data sets from multiple power consumption units, respectively. Finally, a strategy decision optimization network is used to perform multiple rounds of strategy simulation and evaluation on the future supply and consumption prediction data sets to select the candidate allocation strategy with the highest internal evaluation score as the target power allocation strategy.

[0022] In one embodiment of the present invention, multi-dimensional data alignment and time-series coupling analysis are performed on historical power supply data sets and historical power consumption data sets. In a regional power system comprising wind power generation nodes, photovoltaic power generation nodes, and power consumption units in industrial parks and commercial areas, a power data acquisition network acquires historical power supply data sets and historical power consumption data sets at one-hour intervals over the past year. The historical power supply data set includes hourly output data sequences recorded by wind power generation nodes and hourly output data sequences recorded by photovoltaic power generation nodes, while the historical power consumption data set includes hourly electricity load data sequences recorded by industrial park power consumption units and hourly electricity load data sequences recorded by commercial area power consumption units.

[0023] The historical power supply data for each energy node in the historical power supply dataset is cleaned to generate a cleaned supply data sequence. Similarly, the historical power consumption data for each power consumption unit in the historical power consumption dataset is cleaned to generate a cleaned consumption data sequence. In practice, the data cleaning process identifies missing and abnormal data points in both the historical power supply and consumption datasets. For example, the output data for the wind power node at 14:00 on January 15th is missing due to sensor communication interruption, and the electricity load data for the industrial park at 03:00 on July 20th records a negative value due to an malfunction in the acquisition module. For missing data points, data from adjacent time points are used to fill the gaps using time series interpolation. Taking the missing output data for January 15th at 14:00 (0.8 MW) of the aforementioned wind power node as an example, the system uses the output data recorded at January 15th at 13:00 (0.5 MW) and January 15th at 15:00 (0.5 MW) to calculate the filled output data for January 15th at 14:00 using linear interpolation, resulting in 0.65 MW. For abnormal data points, a statistical distribution-based method is used to replace data points exceeding a preset reasonable range with historical power supply or consumption data sets within the corresponding historical period's statistical data. The negative load data of -10 MW recorded in the aforementioned industrial park is identified as abnormal. The preset reasonable range is set to 0 to 50 MW. The system calculates the arithmetic mean of all load data at 03:00 on the third day of July last year to be 28 MW, and uses this average of 28 MW to replace the abnormal -10 MW data point.

[0024] In some embodiments, a unified timestamp index is established. Based on the unified timestamp index, data slices and alignment are performed on the cleaned supply data sequence and the cleaned consumption data sequence to generate time-aligned supply and consumption data. The unified timestamp index uses five minutes as the basic time unit, and the original hourly data is decomposed into twelve consecutive five-minute time unit data points through interpolation. The cleaned wind power output data sequence, photovoltaic power output data sequence, industrial park load data sequence, and commercial area load data sequence are all resampled and aligned to this five-minute timestamp index to form a time-aligned supply and consumption data matrix with consistent length and completely matched time points.

[0025] Time-aligned supply and consumption data are subjected to time-series correlation calculations and mutual information analysis to generate joint supply and consumption data. In practice, the time-series correlation calculation uses Pearson correlation coefficient analysis to analyze the degree of numerical correlation between different sequences at the same time point and across time delays. Mutual information analysis is used to quantify the statistical dependence between supply and consumption data sequences. The calculation process involves labeling the wind power output subsequence in the time-aligned supply and consumption data matrix as a sequence. The photovoltaic power generation output subsequence is labeled as sequence The load subsequence of the industrial park is labeled as a sequence. The commercial area load subsequence is labeled as sequence For any two sequences and their mutual information The calculation is performed using the following formula: in: It is a sequence Values And sequence Values The joint probability, and These are sequences and sequence The marginal probabilities are calculated. By calculating the mutual information values ​​between all supply and consumption sequence pairs, and combining them with the time-series correlation coefficients, joint supply and consumption data are constructed to describe the dynamic relationship between supply and consumption. This joint supply and consumption data can be understood as a multi-dimensional feature tensor, whose dimensions encompass time, combinations of different energy nodes and electricity consumption units, and various correlation metrics.

[0026] In one embodiment of the present invention, see [reference] Figure 2This study performs fluctuation pattern clustering and feature association analysis on joint supply and consumption data, and constructs supply and consumption prediction model configuration tables. Starting from the joint supply and consumption data, it extracts fluctuation feature vectors from the supply data portion and the consumption data portion. The fluctuation feature vectors from the supply data portion are derived from wind power output sequences. and photovoltaic power generation output sequence Extracted from the combined data, the fluctuation feature vector of each supply data segment includes sequences within a specific time window. mean Standard deviation skewness and sequences mean Standard deviation skewness It also includes the Pearson correlation coefficient of the two sequences within this window. The fluctuation feature vector of the consumption data portion comes from the industrial park load sequence. and commercial area load sequence Extracted from the joint data, the fluctuation feature vector of each consumed data portion includes the sequence. mean Standard deviation kurtosis and sequences mean Standard deviation kurtosis and the correlation coefficient between the two sequences. .

[0027] The extracted fluctuation feature vectors from the supply data and consumption data were input into separate clustering analysis engines. The clustering engines used the K-means algorithm with Euclidean distance as the distance metric to perform clustering analysis on the set of fluctuation feature vectors from the supply data. Based on the silhouette coefficient, the fluctuation patterns of the supply data were classified into three categories: high fluctuation and high correlation, low fluctuation and negative correlation, and medium fluctuation and weak correlation. The same clustering analysis was performed on the set of fluctuation feature vectors from the consumption data, classifying the fluctuation patterns into four categories: daytime peak pattern, nighttime trough pattern, holiday smooth pattern, and abnormal peak pattern. Each fluctuation pattern was represented by a central vector. Represents the center vector It contains the average value of all feature vectors in this pattern.

[0028] In some embodiments, the transition probabilities between different volatility patterns are calculated, and historical prediction error data are combined to construct a supply prediction model configuration table and a consumption prediction model configuration table. For the supply data portion, historical data from different patterns are used... Switch to mode frequency Its conversion probability The calculation formula is: in: The supply data section is from the model Convert to mode The probability, Indicates the observed pattern Switch to mode Number of times, Indicates from pattern The total number of transitions from the starting point to all possible modes. The probability of transitions between different modes in the data consumption section. The results were calculated using the same method. Historical forecast error data records the average absolute error values ​​produced by forecasting models using different input variable windows and forecasting step sizes under different supply or consumption fluctuation patterns. The supply forecasting model configuration table is a structured table where each row records a supply fluctuation pattern and its corresponding optimal input variable window. and prediction step size The recommended values, the transition probability matrix between modes, and the historical average prediction error of different model parameters under that mode are recorded in the consumption prediction model configuration table, which records information about consumption fluctuation modes in the same structure.

[0029] Establish a supply-consumption pattern mapping relationship by associating the supply forecasting model configuration table and the consumption forecasting model configuration table. Optionally, the supply-consumption pattern mapping relationship is a joint probability matrix. Matrix elements This indicates that the supply data portion is in a specific fluctuation pattern. At that time, specific fluctuation patterns appeared in the consumed data portion. The joint probability. The value is calculated from the frequency of simultaneous occurrence of the two patterns in historical data. It can be understood that the supply-consumption pattern mapping relationship is input as prior knowledge into the strategy decision optimization network. The strategy generation module in this network queries this mapping relationship when generating the initial power allocation strategy. For example, if the current supply data is partially identified as a high-fluctuation, high-correlation pattern, the strategy generation module will adjust the direction and constraints of strategy generation based on the historical characteristics of consumption patterns with a high joint probability associated with this pattern in the mapping relationship, thereby generating an initial power allocation strategy that better conforms to historical joint fluctuation patterns.

[0030] In one embodiment of the present invention, the optimal input variable window and prediction step size for the energy supply forecasting model and the consumption forecasting model are determined, and the energy supply forecasting model and the consumption forecasting model are run in parallel based on this. Starting from the supply forecasting model configuration table and the consumption forecasting model configuration table, the system searches for the optimal input variable window and prediction step size for the energy supply forecasting model by traversing multiple fluctuation modes in the supply forecasting model configuration table with the goal of minimizing the prediction error. The supply forecasting model configuration table contains three fluctuation modes: high fluctuation and high correlation mode, low fluctuation and negative correlation mode, and medium fluctuation and weak correlation mode. For the high fluctuation and high correlation mode, the system searches within a preset parameter space, which includes the input variable window set {24, 48, 72} hours and the prediction step size set {1, 2, 4} hours. Using historical supply data segments corresponding to the high fluctuation and high correlation mode, the system uses a sliding window method to train a temporary energy supply forecasting model for each combination of input variable window and prediction step size in the parameter space, and calculates the prediction error index on an independent validation dataset. Prediction error index The weighted average absolute error is used for calculation, and the formula is as follows: in: To verify the number of data points, and These are the predicted and actual values ​​of wind power output, respectively. and These are the predicted and actual values ​​of photovoltaic power generation output, respectively. and These are weighting coefficients set based on the installed capacity or historical average output of wind and solar power generation. The system selection makes... The minimum input variable window of 48 hours and the prediction step size of 1 hour are considered the optimal combination for the high volatility and high correlation mode. Using the same method, the low volatility negative correlation mode and the medium volatility weak correlation mode in the supply forecasting model configuration table are traversed to obtain their optimal input variable window and prediction step size, and the search results are stored in the updated supply forecasting model configuration table.

[0031] The optimal input variable window and prediction step size for the consumption prediction model are searched by traversing multiple fluctuation patterns in the consumption prediction model configuration table, with the goal of minimizing the prediction error. The consumption prediction model configuration table contains four fluctuation patterns: daytime peak pattern, nighttime trough pattern, holiday smoothing pattern, and abnormal peak pattern. For the daytime peak pattern, the search is performed within the parameter space of the input variable window set {12, 24, 36} hours and the prediction step size set {0.5, 1, 2} hours. Using historical consumption data for the corresponding pattern, a temporary consumption prediction model is trained for each set of parameters, and the prediction error on the validation set is calculated. . The calculation method and Similarly, but the weighting coefficients apply to the load sequence of the industrial park. and commercial area load sequence The system determines the optimal input variable window as 24 hours and the optimal prediction step size as 0.5 hours for the daytime peak mode. The same process is used to search for optimal parameters for the nighttime off-peak mode, the holiday smooth mode, and the abnormal peak mode, and all results are stored in the updated consumption prediction model configuration table.

[0032] In some embodiments, the current input variable window and prediction step size of the energy supply forecasting model are configured according to the supply forecasting model configuration table, and the supply data of energy nodes acquired in real time are input into the energy supply forecasting model. The system monitors the supply data sequence in real time. and The system calculates the recent fluctuation feature vector and compares it with the center vector of each fluctuation mode in the supply forecasting model configuration table to identify the current supply state as belonging to the "medium fluctuation and weak correlation mode". Subsequently, the system reads the optimal input variable window of 72 hours and the optimal forecast step size of 2 hours corresponding to this mode from the supply forecasting model configuration table and configures the energy supply forecasting model accordingly. The energy supply forecasting model is configured as a deep learning model based on a long short-term memory network, which uses the wind power output sequence of the past 72 hours... and photovoltaic power generation output sequence As input, after internal network calculation, the output is a sequence of predicted values ​​for the next 2 hours, which includes a subset of wind power future supply prediction data and a subset of photovoltaic power future supply prediction data containing the prediction results for four future time points (one point every half hour). Together, they constitute the future supply prediction data set.

[0033] Configure the current input variable window and prediction step size of the consumption prediction model according to the consumption prediction model configuration table, and input the consumption data of the real-time acquired power consumption units into the consumption prediction model. The system monitors the consumption data sequence in real time. and The system identifies the current consumption status as "daytime peak mode." It reads the optimal input variable window of 24 hours and the optimal prediction step size of 0.5 hours corresponding to this mode from the consumption prediction model configuration table, and configures the consumption prediction model accordingly. The consumption prediction model is also a deep learning model based on a long short-term memory network, which calculates the industrial park load sequence over the past 24 hours. and commercial area load sequence As input, the output is the load forecast value for the next 0.5 hours, which constitutes the future consumption forecast dataset. It can be understood that the energy supply forecast model and the consumption forecast model operate in parallel. Based on their respective independent configurations and real-time data streams, they perform calculations simultaneously and output future supply forecast datasets and future consumption forecast datasets respectively.

[0034] See Figure 3 The heatmap presents the distribution of prediction errors under different combinations of input variable windows (24, 48, and 72 hours) and prediction step sizes (1, 2, and 4 hours). The color gradient of the heatmap corresponds to the numerical values, with blue areas representing lower prediction errors and red areas representing higher prediction errors. Specifically, when the input variable window is 48 hours and the prediction step size is 1 hour, the WAE value is 62.8, the minimum among all combinations. This indicates that this parameter combination can most effectively capture the time-series characteristics of supply data under high volatility and high correlation patterns, thereby minimizing prediction errors. Conversely, when the input variable window is 24 hours and the prediction step size is 4 hours, the WAE value reaches 168.3, the maximum among all combinations, reflecting that this parameter combination has poor adaptability to this volatility pattern.

[0035] In one embodiment of the present invention, a strategy decision optimization network is used to perform multiple rounds of strategy simulation and evaluation on a future supply forecast dataset and a future consumption forecast dataset. The strategy decision optimization network is initialized and includes a strategy generation module, a power network simulation module, and a strategy evaluation module. In a microgrid example comprising a wind power node, a photovoltaic power node, an energy storage unit, an industrial power consumption unit, and a commercial power consumption unit, the future supply forecast dataset includes a four-hour wind power output sequence. Megawatt and photovoltaic power generation output sequence Megawatts, future consumption forecast dataset includes industrial load sequences Megawatt and commercial load sequences Megawatts. The strategy generation module generates five initial power allocation strategies based on the future supply forecast data set and the future consumption forecast data set. Each strategy specifies the specific proportion and scheduling instructions for allocating wind power output, photovoltaic power output, and energy storage unit discharge power (positive value) or charging power (negative value) to two power consumption units in each time period within the next four hours.

[0036] For each initial power allocation strategy, the power network simulation module simulates its execution over a preset four-hour period, outputting a simulated supply and consumption trajectory. This trajectory includes the actual power flow of each node within the simulated system, changes in the state of charge of energy storage units, and the supply-demand balance within each time period. The strategy evaluation module analyzes the simulated supply and consumption trajectory based on preset evaluation metrics, calculating an evaluation score for each initial power allocation strategy. These preset metrics include the average supply-demand balance rate. Energy storage unit state of charge health and renewable energy integration rate The strategy evaluation module calculates an evaluation score for each initial power allocation strategy. The calculation formula is as follows: in: It is the average supply and demand balance rate over the entire simulation period. It is the minimum state of charge of the energy storage unit during the simulation cycle (reflecting its health). It is the average renewable energy integration rate over the entire simulation period. , , These are the weighting coefficients assigned to the above indicators, and .

[0037] The strategy decision optimization network sorts the evaluation scores of all initial power allocation strategies and selects one or more initial power allocation strategies with the highest evaluation scores as the elite strategy set. In a specific scenario, the evaluation scores of the five initial strategies are shown in Table 1: Table 1: Initial Power Allocation Strategy Evaluation Score Table Assuming the weighting coefficients are set to After calculating the scores according to the formula, the strategy decision optimization network selects the strategies A and B with the highest scores as the elite strategy set.

[0038] In some embodiments, the strategy generation module generates new power allocation strategies based on an elite strategy set through strategy crossover and strategy mutation operations, forming a new generation strategy set. The strategy crossover operation combines the allocation schemes for the first two time periods of Strategy A with the allocation schemes for the last two time periods of Strategy B to generate a new strategy. The strategy mutation operation randomly changes the energy storage unit charging and discharging power command for a certain time period in the elite strategy; for example, it changes the energy storage charging power for the second time period in Strategy A from -0.2 MW to -0.1 MW, generating another new strategy. Through crossover and mutation, the strategy generation module generates five new power allocation strategies. The strategy decision optimization network repeatedly executes the power network simulation and strategy evaluation process for each new power allocation strategy in the new generation strategy set, calculating a new evaluation score.

[0039] See Figure 4 This study presents the temporal dynamics of wind and solar power supply and industrial and commercial power consumption in a microgrid scenario over the next four hours. On the supply side, solar power output shows a "rise then fall" trend, reaching a peak of 3.2MW in the second hour before gradually declining, a fluctuation consistent with typical daytime solar power output patterns. Wind power output, on the other hand, exhibits a "fluctuating rise" characteristic, gradually increasing from 1.2MW in the first hour to 1.6MW in the fourth hour, demonstrating relatively better overall supply stability. On the consumption side, industrial load shows a "fluctuating growth" trend, experiencing a brief dip in the third hour before rebounding to a peak of 3.0MW in the fourth hour, reflecting the intermittent yet continuous electricity consumption characteristics of industrial production. Commercial load, however, shows a "steady rise" trend, gradually increasing from 1.2MW in the first hour to 1.5MW in the fourth hour, aligning with the electricity consumption patterns of commercial activities. The time-series dynamics of supply and consumption presented in the figure are the core input data for the strategy decision optimization network to generate the initial power allocation strategy. Its fluctuation characteristics will directly affect the calculation of key evaluation indicators such as supply and demand balance rate, energy storage charge state health and renewable energy consumption rate in subsequent rounds of strategy simulation.

[0040] In one embodiment of the invention, the target power allocation strategy is executed in an actual power network, collecting actual supply and consumption data generated during the execution process. The target power allocation strategy ultimately selects candidate allocation strategies, which specify detailed scheduling instructions for each time period within the next four hours. In a microgrid example, the system controls energy storage units, adjusts the output of distributed power sources, and supplies power to industrial and commercial power consumption units according to the strategy instructions. Throughout the four-hour execution cycle, the power data acquisition network synchronously collects the actual output sequence of wind power generation nodes. Actual output sequence of megawatt and photovoltaic power generation nodes Actual load sequence of megawatts and industrial power consumption units Actual load sequence of megawatts and commercial electricity consumption units Megawatts.

[0041] This calculation determines the supply forecast bias between actual supply data and the future supply forecast dataset, as well as the consumption forecast bias between actual consumption data and the future consumption forecast dataset. The future supply forecast dataset is derived from the output of the energy supply forecast model and includes a wind power generation forecast sequence. Megawatt and photovoltaic power generation forecast series Megawatts. The output of the future consumption forecasting model, containing industrial load forecast sequences. Megawatt and Commercial Load Forecast Sequence Megawatts. Supply forecast bias is the set of differences between predicted and actual values ​​at each point in time. For example, the supply forecast bias sequence for wind power generation is... Megawatts. Consumption forecast bias is also the set of differences between predicted and actual values ​​at each point in time. For example, the consumption forecast bias sequence for commercial load is... Megawatts.

[0042] Based on supply and consumption forecast biases, the internal parameters of the energy supply and consumption forecast models are adjusted in reverse. Both the energy supply and consumption forecast models are deep learning models based on Long Short-Term Memory (LSTM) networks. Supply forecast bias sequence. and Combined into a comprehensive supply deviation indicator It is used to guide the parameter updates of energy supply forecasting models. The calculation formula is: in: This is the predicted total duration (4 in this case). and They are time points The predicted and actual output of wind power generation and They are time points The predicted and actual output of photovoltaic power generation. The consumption prediction deviation is calculated using the same method to obtain the comprehensive consumption deviation index. The reverse adjustment process is based on... and As part of the loss signal, the backpropagation algorithm is used to update the weight matrix and bias vector in the long short-term memory network units of the energy supply prediction model and the consumption prediction model. Optionally, the update process uses a momentum-driven stochastic gradient descent optimizer with a learning rate of 0.001 and a momentum parameter of 0.9.

[0043] The actual performance of the target power allocation strategy is compared with the evaluation score given by the strategy decision optimization network in the simulation. Based on the comparison results, the weights of the evaluation indicators in the strategy evaluation module are adjusted. The evaluation score given by the strategy decision optimization network for the target power allocation strategy in the simulation. The value is 0.88. The actual implementation effect needs to be quantified by calculating the average supply-demand balance rate over the actual implementation period. Energy storage unit state of charge health (Minimum state of charge value within the period) and actual renewable energy integration rate Calculated based on actual data , , According to the original weighting coefficients of the strategy evaluation module ( ) Calculate the overall score of actual effect .

[0044] See Figure 5 During model calibration, the comparison experiment of LSTM model parameter adjustment magnitude was performed using a momentum-driven stochastic gradient descent (SGD) optimizer, with a learning rate of 0.001 and a momentum parameter of 0.9. The experiment quantified the parameter update magnitude during backpropagation for six types of parameters: input layer weights, hidden layer weights, output layer weights, input layer biases, hidden layer biases, and output layer biases, for both the supply prediction model (blue bars) and the consumption prediction model (orange bars). The results show that the parameter adjustment of the supply prediction model is predominantly negative, with the output layer weights showing the largest negative adjustment magnitude (approximately -0.003), followed by the input layer weights and hidden layer biases, at approximately -0.0025 and -0.0015 respectively. Conversely, the hidden layer weights and output layer biases show positive updates, with magnitudes of approximately 0.0018 and 0.0021 respectively. The parameter adjustments of the consumption prediction model exhibit significant bidirectional fluctuations: the positive adjustment magnitude of the output layer weights reaches a peak (approximately 0.0035), while the positive adjustments of the hidden layer bias and input layer weights are approximately 0.002 and 0.0019, respectively; simultaneously, the negative adjustments of the hidden layer weights and output layer bias are quite prominent, at approximately -0.003 and -0.0025, respectively. This difference in parameter adjustments reflects the inherent differences in data fluctuation patterns and feature correlation strength between supply and consumption prediction tasks, and also verifies the effectiveness of the dual-model parallel calibration strategy in adapting to the characteristics of different prediction tasks.

[0045] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A deep learning-based method for optimizing and managing new energy power allocation strategies, characterized in that, The method includes: Through the power data acquisition network, acquire historical power supply data sets of multiple energy nodes and historical power consumption data sets of multiple power consumption units within a preset historical period; Multidimensional data alignment and time-series coupling analysis are performed on the historical power supply data set and the historical power consumption data set to generate joint supply and consumption data; In the deep prediction model construction space, fluctuation pattern clustering and feature correlation analysis are performed on the joint supply and consumption data to determine the optimal input variable window and prediction step size for the energy supply prediction model and the consumption prediction model. Based on the optimal input variable window and prediction step size, the energy supply prediction model and the consumption prediction model are run in parallel to obtain future supply prediction data sets for multiple energy nodes and future consumption prediction data sets for multiple power consumption units, respectively. Using a strategy decision optimization network, multiple rounds of strategy simulation and evaluation are performed on the future supply forecast data set and the future consumption forecast data set. The candidate allocation strategy with the highest evaluation score within the strategy decision optimization network is selected as the target power allocation strategy.

2. The deep learning-based new energy power allocation strategy decision optimization management method as described in claim 1, characterized in that, The multidimensional data alignment and time-series coupling analysis of the historical power supply data set and the historical power consumption data set includes: Data cleaning is performed on the historical power supply data of each energy node in the historical power supply data set to generate a cleaned supply data sequence. Data cleaning is performed on the historical power consumption data of each power consumption unit in the historical power consumption dataset to generate a cleaned consumption data sequence. Establish a unified timestamp index, and perform data slicing and alignment on the cleaned supply data sequence and the cleaned consumption data sequence based on the unified timestamp index to generate time-aligned supply and consumption data; The time-aligned supply and consumption data are subjected to time-series correlation calculation and mutual information analysis to generate the joint supply and consumption data.

3. The deep learning-based new energy power allocation strategy decision optimization management method as described in claim 2, characterized in that, The process of performing fluctuation pattern clustering and feature correlation analysis on the joint supply and consumption data within the deep prediction model construction space includes: Extract the fluctuation feature vectors of the supply data portion and the consumption data portion from the joint supply and consumption data; The fluctuation feature vectors of the supply data portion and the fluctuation feature vectors of the consumption data portion are respectively input into independent clustering analysis engines; The clustering analysis engine performs clustering analysis on the fluctuation feature vectors of the supply data portion and the consumption data portion respectively, and identifies multiple fluctuation patterns. Calculate the conversion probability between different fluctuation patterns, and construct supply forecast model configuration tables and consumption forecast model configuration tables by combining historical forecast error data.

4. The deep learning-based new energy power allocation strategy decision optimization management method as described in claim 3, characterized in that, The determination of the optimal input variable window and prediction step size for the energy supply prediction model and consumption prediction model includes: By iterating through the various fluctuation patterns in the supply forecasting model configuration table, with the goal of minimizing the forecasting error, the optimal input variable window and forecasting step size of the energy supply forecasting model are searched. The system iterates through various fluctuation patterns in the consumption prediction model configuration table, searching for the optimal input variable window and prediction step size of the consumption prediction model with the goal of minimizing the prediction error. The optimal input variable window and prediction step size of the energy supply prediction model and the optimal input variable window and prediction step size of the consumption prediction model obtained from the search are stored in the supply prediction model configuration table and the consumption prediction model configuration table, respectively.

5. The deep learning-based new energy power allocation strategy decision optimization management method as described in claim 4, characterized in that, The parallel operation of the energy supply prediction model and the consumption prediction model includes: According to the supply forecasting model configuration table, configure the current input variable window and forecasting step size of the energy supply forecasting model, and input the supply data of the energy nodes acquired in real time into the energy supply forecasting model. According to the consumption prediction model configuration table, configure the current input variable window and prediction step size of the consumption prediction model, and input the consumption data of the real-time power consumption unit into the consumption prediction model. The energy supply forecasting model outputs the set of future supply forecast data based on its input supply data. The consumption prediction model outputs the set of future consumption prediction data based on its input consumption data.

6. The deep learning-based new energy power allocation strategy decision optimization management method as described in claim 1, characterized in that, The process of utilizing a strategy decision optimization network to perform multiple rounds of strategy simulation and evaluation on the future supply forecast data set and the future consumption forecast data set includes: Initialize the policy decision optimization network, which includes a policy generation module, a power network simulation module, and a policy evaluation module; The strategy generation module generates multiple initial power allocation strategies based on the future supply forecast data set and the future consumption forecast data set. For each initial power allocation strategy, the power network simulation module simulates its execution process within a preset future time period and outputs the simulated supply and consumption trajectory. The strategy evaluation module analyzes the simulated supply and consumption trajectory based on preset evaluation indicators and calculates the evaluation score for each initial power allocation strategy.

7. The deep learning-based new energy power allocation strategy decision optimization management method as described in claim 6, characterized in that, The candidate allocation strategies that have the highest internal evaluation scores in the strategy decision optimization network include: The strategy decision optimization network sorts the evaluation scores of all initial power allocation strategies and selects one or more initial power allocation strategies with the highest evaluation scores as the elite strategy set. The strategy generation module generates new power allocation strategies based on the elite strategy set through strategy crossover and strategy mutation operations, thus forming a new generation of strategy set; For each new power allocation strategy in the next-generation strategy set, the strategy decision optimization network repeatedly executes the power network simulation and the strategy evaluation process to calculate a new evaluation score. The process of strategy evaluation, elite strategy selection, strategy generation and simulation is executed iteratively until the preset iteration termination condition is reached. The strategy with the highest evaluation score in the elite strategy set after the last iteration is determined as the candidate allocation strategy.

8. The deep learning-based new energy power allocation strategy decision optimization management method as described in claim 7, characterized in that, The method also includes model refinement and parameter calibration of the policy decision optimization network: The target power allocation strategy is executed in the actual power network, and the actual supply data and actual consumption data generated during the actual execution process are collected. Calculate the supply forecast deviation between the actual supply data and the future supply forecast data set, and the consumption forecast deviation between the actual consumption data and the future consumption forecast data set; Based on the supply forecast deviation and the consumption forecast deviation, the internal parameters of the energy supply forecast model and the consumption forecast model are adjusted in reverse. The actual execution effect of the target power allocation strategy is compared with the evaluation score given by the strategy decision optimization network in the simulation, and the evaluation index weights of the strategy evaluation module are adjusted according to the comparison results.

9. The deep learning-based new energy power allocation strategy decision optimization management method as described in claim 2, characterized in that, The data cleaning includes: Identify missing and abnormal data points in the historical power supply data set and the historical power consumption data set; For missing data points, data from time points adjacent to the missing data points are used to fill them using time series interpolation methods; For abnormal data points, a statistical distribution-based method is used to replace data points that exceed a preset reasonable range with statistical data from the historical power supply data set or the historical power consumption data set in the corresponding historical period.

10. The deep learning-based new energy power allocation strategy decision optimization management method as described in claim 3, characterized in that, The configuration tables for constructing supply forecasting models and consumption forecasting models also include: Associate the supply forecasting model configuration table and the consumption forecasting model configuration table to establish a supply-consumption pattern mapping relationship; The supply-consumption pattern mapping relationship is used to record the joint probability that the consumption data part will exhibit a specific fluctuation pattern when the supply data part is in a specific fluctuation pattern. The supply and consumption pattern mapping relationship is used as prior knowledge and input into the strategy decision optimization network to guide the strategy generation module in generating an initial power allocation strategy that is more consistent with historical joint fluctuation patterns.

Images