Data center hybrid energy storage and renewable energy source intelligent scheduling control method and system

Abstract

The invention provides a data center hybrid energy storage and renewable energy intelligent scheduling control method and system, and relates to the technical field of data energy storage and resource scheduling, and the method comprises the steps: carrying out the load prediction through a double-layer time sequence memory network in combination with a multi-head attention mechanism, building a hybrid energy storage optimization scheduling model based on a renewable energy generation power prediction result, and carrying out the optimization scheduling of the hybrid energy storage. And a double-layer model predictive control framework is adopted to output a corrected charging and discharging power instruction, so that optimal control of the hybrid energy storage system is realized, the energy utilization efficiency is improved, and the operation cost of a data center is reduced.

Description

Technical Field

[0001] The present invention relates to data energy storage and resource scheduling technology, and in particular to a data center hybrid energy storage and renewable energy intelligent scheduling control method and system. Background Art

[0002] With the rapid development of cloud computing and internet technologies, data centers have become essential infrastructure for the modern information society. However, the high energy consumption of data centers is becoming increasingly prominent, and their electricity demand continues to grow, placing enormous pressure on the power grid. To reduce traditional energy consumption and carbon emissions, an increasing number of data centers are integrating renewable energy and hybrid energy storage systems to improve energy efficiency and system stability.

[0003] Data center electricity loads are subject to significant volatility and uncertainty, while renewable energy sources such as photovoltaics and wind power also exhibit random characteristics influenced by meteorological conditions. Traditional data center energy management technologies rely primarily on simple historical statistical data or basic linear forecasting methods to forecast electricity load and renewable energy generation. These methods fail to effectively capture complex time series characteristics and the influence of external factors, resulting in low forecast accuracy and compromising the scientific nature of subsequent scheduling decisions.

[0004] Existing hybrid energy storage control methods often use a single optimization algorithm, which makes it difficult to balance the differences in characteristics of different types of energy storage devices in the energy storage system, such as the high power density characteristics of supercapacitors and the high energy density characteristics of lithium batteries. This makes it impossible to achieve the optimal configuration of energy storage resources, reducing the economy and reliability of the system.

[0005] In addition, most energy storage scheduling and control strategies in existing technologies lack an effective response mechanism to the uncertainty of system operation. When there is a large deviation between the actual operating data and the predicted data, the control effect is significantly reduced, making it difficult to ensure the stable operation and optimal energy efficiency of the system. This is especially true when large-scale renewable energy is connected and load changes occur suddenly. Summary of the Invention

[0006] The embodiments of the present invention provide a data center hybrid energy storage and renewable energy intelligent scheduling control method and system, which can solve the problems in the prior art.

[0007] A first aspect of an embodiment of the present invention provides a method for intelligent scheduling and control of hybrid energy storage and renewable energy in a data center, comprising:

[0008] Collect data center power load data, hybrid energy storage system operating status data, and renewable energy power generation data;

[0009] A two-layer temporal memory network is used to extract the temporal features of the power load data, and the temporal features are processed through a multi-head attention mechanism to establish a data center power load forecasting model. The power load of the data center in a future time period is predicted based on the power load forecasting model to obtain predicted power load data;

[0010] Establishing a renewable energy power generation prediction model based on the renewable energy power generation data, and predicting the renewable energy power generation in a future time period according to the renewable energy power generation prediction model to obtain predicted power generation data;

[0011] Constructing a hybrid energy storage optimization scheduling model after feature enhancement based on the predicted electricity load data and the predicted power generation data, and outputting initial charge and discharge power instructions based on the hybrid energy storage optimization scheduling model;

[0012] Based on the operating status data of the hybrid energy storage system and the initial charge and discharge power instructions, a two-layer model predictive control framework is used for optimization, and a corrected charge and discharge power instruction is output through the synergy of distributed optimization and a robust controller;

[0013] The charge and discharge operations of the hybrid energy storage system are controlled according to the corrected charge and discharge power instructions.

[0014] A two-layer temporal memory network is used to extract the temporal features of the power load data. The temporal features are processed through a multi-head attention mechanism to establish a data center power load forecasting model. The power load of the data center in the future time period is predicted based on the power load forecasting model to obtain predicted power load data, including:

[0015] Preprocessing the power load data, and converting the power load data into standardized load data using a maximum and minimum value normalization method;

[0016] A double-layer temporal memory network is used to extract the temporal features of the standardized load data, a random dropout layer is set between the two layers of the double-layer temporal memory network, and the temporal features are fused with the time features, the temperature features, and the load features to obtain fused features;

[0017] Processing the fused features through a multi-head attention mechanism to calculate a multi-head attention weight, where the multi-head attention weight is determined by the attention scores calculated by the multiple attention heads, and the attention scores are calculated by the query vector, the key vector, and the value vector;

[0018] Performing a weighted operation on the multi-head attention weights and the fusion features to obtain an attention feature, and establishing a data center power load prediction model based on the attention feature;

[0019] Constructing a loss function including a weighted combination of a root mean square error term and a mean absolute percentage error term, and optimizing and training the data center power load forecasting model based on the loss function;

[0020] Inputting the characteristic data corresponding to the time period to be predicted into the trained data center power load prediction model to obtain an initial prediction result;

[0021] The confidence interval of the initial prediction result is calculated based on the historical prediction error distribution, and when the initial prediction result exceeds the confidence interval, correction is performed to obtain predicted electricity load data.

[0022] Establishing a renewable energy power generation prediction model based on the renewable energy power generation data, and predicting the renewable energy power generation in a future time period according to the renewable energy power generation prediction model to obtain predicted power generation data, including:

[0023] Preprocessing the power generation data of the renewable energy, performing noise reduction processing on the power generation data through wavelet transform, and converting the noise-reduced power generation data into standardized power generation data;

[0024] Constructing a hybrid prediction model, wherein the period decomposition unit of the hybrid prediction model uses a variational mode decomposition method to decompose the standardized power generation data into a trend component and a fluctuation component, the trend prediction unit of the hybrid prediction model uses a long short-term memory network structure to model the trend component, and the fluctuation prediction unit of the hybrid prediction model uses a gated recurrent neural network structure to model the fluctuation component;

[0025] The meteorological forecast data is input into the hybrid forecast model, and the hybrid forecast model is optimized and trained based on the combined loss function of the root mean square error and the mean absolute error. The meteorological forecast data corresponding to the time period to be predicted is input into the trained hybrid forecast model to obtain the predicted power generation data.

[0026] After feature enhancement is performed on the predicted power load data and the predicted power generation data, a hybrid energy storage optimization scheduling model is constructed, and initial charge and discharge power instructions are output according to the hybrid energy storage optimization scheduling model, including:

[0027] constructing a multidimensional spatiotemporal state space based on the predicted electricity load data and the predicted power generation data, and performing feature enhancement on the multidimensional spatiotemporal state space using a spatiotemporal attention mechanism to obtain an enhanced state space;

[0028] Constructing a hierarchical deep reinforcement learning network, wherein the state encoder of the hierarchical deep reinforcement learning network uses a dual-stream graph convolutional neural network to respectively extract the temporal features and spatial features of the enhanced state space; the action generator of the hierarchical deep reinforcement learning network fuses the temporal features with the spatial features based on a gated recurrent unit to generate a charge-discharge power decision; and the value evaluator of the hierarchical deep reinforcement learning network uses a dual evaluation network structure to evaluate the value function of the charge-discharge power decision;

[0029] Based on historical scheduling data, an immediate reward term containing multiple costs is constructed. A recurrent neural network is used to model the charging and discharging operation sequence to obtain a long-term reward term. The immediate reward term is dynamically adjusted through adaptive weight optimization to obtain an adaptive hybrid reward function.

[0030] Combining the adaptive hybrid reward function with the hierarchical deep reinforcement learning network to construct a hybrid energy storage optimization scheduling model, and training the hybrid energy storage optimization scheduling model;

[0031] The current state information is input into the trained hybrid energy storage optimization scheduling model, and the initial charge and discharge power instructions are output through the action generator.

[0032] Based on historical scheduling data, an immediate reward term containing multiple costs is constructed. A recurrent neural network is used to model the charging and discharging operation sequence to obtain a long-term reward term. The immediate reward term is dynamically adjusted through adaptive weight optimization. The combination results in an adaptive hybrid reward function, including:

[0033] Collecting historical dispatch data of the hybrid energy storage system, and constructing an instant reward item including multiple costs based on the historical dispatch data;

[0034] Using a recurrent neural network to perform time series modeling on the charge and discharge operation sequences in the historical scheduling data, extracting long-term dependency features of the charge and discharge operation sequences, and constructing a long-term reward item reflecting the long-term operating benefits of the energy storage system based on the long-term dependency features;

[0035] An evaluation index is constructed based on the system operation state sequence and the cost-benefit sequence. A scheduling effect score of each group of charging and discharging operations in the historical scheduling data is calculated based on the evaluation index. The scheduling effect score is input into an adaptive weight optimizer. The adaptive weight optimizer dynamically adjusts the weight coefficient of each cost sub-item in the instant reward item according to the changing trend of the scheduling effect score. When the scheduling effect score corresponding to a specific cost sub-item decreases, the weight coefficient of the cost sub-item is increased;

[0036] The immediate reward item and the long-term reward item are weightedly combined to obtain an adaptive hybrid reward function.

[0037] Based on the operating status data of the hybrid energy storage system and the initial charge and discharge power instructions, a two-layer model predictive control framework is used for optimization, and a corrected charge and discharge power instruction is output through the synergy of distributed optimization and a robust controller, including:

[0038] According to the operating status data of the hybrid energy storage system, the energy storage units are grouped to construct a hierarchical system state space, state feature vectors are extracted based on a variational autoencoder, and spatiotemporal features of the state feature vectors are extracted using a hybrid structure of a gated recurrent neural network and a residual convolutional network to construct a multimodal state prediction model;

[0039] Establish a system dynamic response model, use the output of the multimodal state prediction model to perform state estimation through a particle filter algorithm, perform probability distribution fitting on the prediction results of the system dynamic response model, and construct an adaptive constraint evaluator based on the information entropy criterion;

[0040] Constructing a two-layer model predictive control framework, wherein the long-term optimization layer performs predictive optimization based on the multimodal state prediction model, and the short-term execution layer dynamically corrects the initial charge and discharge power instructions according to the prediction error of the long-term optimization layer to obtain an initial correction result;

[0041] Constructing a multi-time-scale optimization objective function based on the initial correction result, decomposing the multi-time-scale optimization objective function using a distributed alternating direction multiplier method to solve subproblems, and inputting the solution results into a robust controller;

[0042] The initial charge and discharge power instructions are collaboratively optimized with the robust controller through the two-layer model predictive control framework to output a corrected charge and discharge power instruction.

[0043] A two-layer model predictive control framework is constructed, wherein the long-term optimization layer performs predictive optimization based on the multimodal state prediction model, and the short-term execution layer dynamically corrects the initial charge and discharge power instructions according to the prediction error of the long-term optimization layer to obtain an initial correction result, including:

[0044] Constructing a multimodal state space, and mapping the multimodal state space into a long-term prediction state set and a short-term execution state set;

[0045] Based on the long-term prediction state set, a recurrent neural network structure with gated skip connections is used to extract multi-scale state time series features, the multi-scale state time series features are input into a prediction optimizer constructed by a conditional generative adversarial network, and a prediction state sequence is output;

[0046] Establishing a short-term execution layer based on the short-term execution state set, wherein the short-term execution layer uses a bidirectional long short-term memory network to calculate the temporal correlation between the predicted state sequence and the actual state sequence to generate a prediction error feature matrix;

[0047] Constructing an adaptive corrector, wherein the adaptive corrector performs multi-dimensional feature extraction on the prediction error feature matrix to obtain multi-dimensional features, and constructs a multi-head cross attention mechanism based on the multi-dimensional features to generate adaptive correction weights;

[0048] A deep neural network with residual connections is used to perform nonlinear feature fusion on the adaptive correction weight, the multi-scale state time series feature, and the initial charge and discharge power instruction to obtain a nonlinear feature fusion result;

[0049] The nonlinear feature fusion result is input into a probability distribution generation network. The probability distribution generation network samples and generates a set of candidate correction coefficients in the latent variable space based on the distribution characteristics of the prediction error feature matrix, evaluates the set of candidate correction coefficients, selects the optimal correction coefficient based on the Bayesian optimization criterion, and dynamically adjusts the initial charge and discharge power instructions using the optimal correction coefficient to obtain an initial correction result.

[0050] A second aspect of an embodiment of the present invention provides a data center hybrid energy storage and renewable energy intelligent dispatching and control system, including:

[0051] The first unit is used to collect data on the power load of the data center, the operating status data of the hybrid energy storage system, and the power generation data of renewable energy;

[0052] The second unit is configured to extract time series features of the power load data using a two-layer temporal memory network, process the time series features using a multi-head attention mechanism to establish a data center power load forecasting model, and predict the data center power load in a future time period based on the power load forecasting model to obtain predicted power load data;

[0053] A third unit is configured to establish a renewable energy power generation prediction model based on the renewable energy power generation data, and predict the renewable energy power generation in a future time period according to the renewable energy power generation prediction model to obtain predicted power generation data;

[0054] A fourth unit is configured to construct a hybrid energy storage optimization scheduling model after feature enhancement based on the predicted electricity load data and the predicted power generation data, and output an initial charge and discharge power instruction based on the hybrid energy storage optimization scheduling model;

[0055] a fifth unit, configured to optimize the hybrid energy storage system based on the operating status data and the initial charge and discharge power instructions using a two-layer model predictive control framework, and output a corrected charge and discharge power instruction through the synergy of distributed optimization and a robust controller;

[0056] The sixth unit is configured to control the charge and discharge operations of the hybrid energy storage system according to the corrected charge and discharge power instructions.

[0057] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0058] processor;

[0059] a memory for storing processor-executable instructions;

[0060] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0061] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0062] The beneficial effects of this application are as follows:

[0063] The data center hybrid energy storage and renewable energy intelligent scheduling and control method provided by the present invention extracts timing features through a double-layer temporal memory network and combines it with a multi-head attention mechanism to construct a load forecasting model. It can accurately predict the power load of the data center, improve the prediction accuracy, and lay the foundation for subsequent energy storage scheduling.

[0064] The hybrid energy storage optimization scheduling model established in this invention performs feature enhancement processing on the predicted data, comprehensively considers the characteristics of renewable energy power generation and the electricity demand of the data center, and proposes a distributed optimization and robust control scheme based on a two-layer model predictive control framework, which effectively solves the system uncertainty problem and improves the operational stability of the energy storage system.

[0065] The present invention achieves efficient utilization of renewable energy and a stable supply of data center electricity demand through precise charge and discharge control of the hybrid energy storage system, reducing data center operating costs, improving system energy efficiency, and reducing dependence on the power grid, with significant economic and environmental benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 Schematic diagram of the process of intelligent scheduling and control method of hybrid energy storage and renewable energy for data centers according to an embodiment of the present invention;

[0067] Figure 2This is a bar chart comparing the performance of the data center power load prediction model according to an embodiment of the present invention;

[0068] Figure 3 Schematic diagram of the comparison of root mean square errors of different prediction models according to the embodiment of the present invention;

[0069] Figure 4 Constructing a flow chart for the adaptive hybrid reward function according to an embodiment of the present invention;

[0070] Figure 5 This is a bar chart comparing the performance of the two-layer model predictive control framework according to the embodiment of the present invention. DETAILED DESCRIPTION

[0071] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0072] The technical solution of the present invention is described in detail below with reference to specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0073] Figure 1 FIG. 1 is a flow chart of a method for intelligent scheduling and control of hybrid energy storage and renewable energy in a data center according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0074] Collect data center power load data, hybrid energy storage system operating status data, and renewable energy power generation data;

[0075] A two-layer temporal memory network is used to extract the temporal features of the power load data, and the temporal features are processed through a multi-head attention mechanism to establish a data center power load forecasting model. The power load of the data center in a future time period is predicted based on the power load forecasting model to obtain predicted power load data;

[0076] Establishing a renewable energy power generation prediction model based on the renewable energy power generation data, and predicting the renewable energy power generation in a future time period according to the renewable energy power generation prediction model to obtain predicted power generation data;

[0077] Constructing a hybrid energy storage optimization scheduling model after feature enhancement based on the predicted electricity load data and the predicted power generation data, and outputting initial charge and discharge power instructions based on the hybrid energy storage optimization scheduling model;

[0078] Based on the operating status data of the hybrid energy storage system and the initial charge and discharge power instructions, a two-layer model predictive control framework is used for optimization, and a corrected charge and discharge power instruction is output through the synergy of distributed optimization and a robust controller;

[0079] The charge and discharge operations of the hybrid energy storage system are controlled according to the corrected charge and discharge power instructions.

[0080] In an optional embodiment, a two-layer temporal memory network is used to extract the temporal features of the power load data, and the temporal features are processed through a multi-head attention mechanism to establish a data center power load forecasting model. The data center power load in a future time period is predicted based on the power load forecasting model to obtain predicted power load data, including:

[0081] Preprocessing the power load data, and converting the power load data into standardized load data using a maximum and minimum value normalization method;

[0082] A double-layer temporal memory network is used to extract the temporal features of the standardized load data, a random dropout layer is set between the two layers of the double-layer temporal memory network, and the temporal features are fused with the time features, the temperature features, and the load features to obtain fused features;

[0083] Processing the fused features through a multi-head attention mechanism to calculate a multi-head attention weight, where the multi-head attention weight is determined by the attention scores calculated by the multiple attention heads, and the attention scores are calculated by the query vector, the key vector, and the value vector;

[0084] Performing a weighted operation on the multi-head attention weights and the fusion features to obtain an attention feature, and establishing a data center power load prediction model based on the attention feature;

[0085] Constructing a loss function including a weighted combination of a root mean square error term and a mean absolute percentage error term, and optimizing and training the data center power load forecasting model based on the loss function;

[0086] Inputting the characteristic data corresponding to the time period to be predicted into the trained data center power load prediction model to obtain an initial prediction result;

[0087] The confidence interval of the initial prediction result is calculated based on the historical prediction error distribution, and when the initial prediction result exceeds the confidence interval, correction is performed to obtain predicted electricity load data.

[0088] Obtain historical data center power load data, including hourly power consumption data for the past 30 days. Preprocess the acquired power load data and convert it into standardized load data using the maximum and minimum normalization method. Specifically, for the original load data x, convert it to standardized data x' using a formula that maps all data to the interval [0, 1]. For example, if the maximum power load in the original data set is 10,000 kW and the minimum power load is 2,000 kW, then the normalized value for a power load of 8,000 kW is 0.75.

[0089] A two-layer temporal memory network (TMN) was used to extract time series features from the standardized load data. This network consists of two recurrent neural networks, each containing 128 hidden units. The output of the first layer serves as the input to the second layer. A random dropout layer with a dropout rate of 0.2 is placed between the two layers to prevent overfitting. The input data is the electricity load sequence for the past 168 hours (7 days), and a sliding window mechanism is used to construct training samples. After network processing, a 128-dimensional time series feature vector is obtained.

[0090] At the same time, auxiliary features related to power load are extracted, including: time features (including time information such as hour, date, week, month, and season, encoded as a 24-dimensional vector); temperature features (including current temperature, maximum temperature, and minimum temperature, a total of 3 dimensions); and load features (including server CPU utilization, memory usage, network traffic, etc., a total of 12 dimensions). These auxiliary features are spliced ​​with the time series features to form a 167-dimensional fused feature vector.

[0091] The fused features are processed by a multi-head attention mechanism to calculate the attention weight. In this embodiment, 8 attention heads are used, and the hidden dimension of each attention head is 64. For each attention head, the fused features are linearly transformed into a query vector, a key vector, and a value vector. The attention score is obtained by performing a dot product operation on the query vector and the key vector, and after scaling and softmax normalization. The attention score is weighted and summed with the value vector to obtain the output of the attention head. Finally, the outputs of the 8 attention heads are spliced ​​and linearly transformed to obtain the output of the multi-head attention with a dimension of 512. This mechanism can capture the dependencies between different time steps and improve the model's ability to model long-term dependencies.

[0092] The output of the multi-head attention is processed through a feedforward neural network consisting of two fully connected layers. The first layer contains 256 neurons and uses the ReLU activation function. The second layer outputs a dimension of 1, corresponding to the predicted power load value. This completes the construction of the data center power load forecasting model.

[0093] To optimize model performance, a loss function is constructed that includes a weighted combination of the root mean square error (RMSE) and mean absolute percentage error (MAPE). In this example, the RMSE weight is set to 0.6, and the MAPE weight is set to 0.4. This combined loss function considers both the absolute and relative errors between the predicted and true values, enabling the model to maintain good prediction accuracy under both high and low load conditions.

[0094] Model training uses mini-batch gradient descent with a batch size of 64. The initial learning rate is set to 0.001, and a learning rate decay strategy is adopted, with the learning rate multiplied by 0.9 every 50 training cycles. Training lasts for 300 cycles or is stopped early when the loss on the validation set does not improve for 20 consecutive cycles.

[0095] After the model training is completed, for the electricity load to be predicted in the next 24 hours, the characteristic data of the corresponding time period (including time characteristics, predicted temperature characteristics, estimated load characteristics, etc.) are input into the trained model to obtain the initial prediction results.

[0096] A confidence interval is calculated for the initial forecast result based on the historical forecast error distribution. Specifically, the forecast errors for the same period over the past 30 days are tallied, and their mean μ and standard deviation σ are calculated. A 95% confidence interval is defined as μ ± 2σ. If the initial forecast result falls outside this confidence interval, the forecast value is revised to the bounds of the confidence interval. For example, if the forecast load at a certain moment is 9500kW, but the upper limit of the confidence interval is 9200kW, the forecast result is revised to 9200kW.

[0097] After completing these steps, we ultimately obtain predicted power load data for the future time period. In practical applications, this method achieved excellent performance on a test set, with a mean absolute percentage error of 2.3% and a root mean square error of 185kW. This improves prediction accuracy by approximately 25% compared to traditional methods, effectively supporting data center energy planning and optimization.

[0098] Figure 2 This is a bar chart comparing the performance of the data center power load prediction model according to an embodiment of the present invention:

[0099] This chart compares the performance of different models across multiple evaluation metrics, including three model types: a two-layer temporal memory network model (blank bars), a multi-head attention model (shaded bars), and a fusion model (checkered bars). Looking at four key metrics: In terms of prediction accuracy, the fusion model performed best, reaching 96.8%, slightly higher than the 94.2% of the multi-head attention model and 92.7% of the two-layer temporal model. In terms of peak load prediction, the fusion model also led, reaching 94.1%, compared to 89.5% and 87.2% for the multi-head attention model and the two-layer temporal model, respectively. In terms of 24-hour prediction error, all three models maintained low levels, with the fusion model having the lowest error of 2.5%, the multi-head attention model at 4.3%, and the two-layer temporal model at 5.2%. In terms of abnormal load responsiveness, the fusion model continued to perform best, reaching 92.7%, compared to 87.2% for the multi-head attention model and 83.1% for the two-layer temporal model. Overall, the fusion model demonstrated optimal performance across all evaluation metrics.

[0100] In an optional embodiment, a renewable energy power generation prediction model is established based on the renewable energy power generation data, and the renewable energy power generation in a future time period is predicted according to the renewable energy power generation prediction model to obtain the predicted power generation data, including:

[0101] Preprocessing the power generation data of the renewable energy, performing noise reduction processing on the power generation data through wavelet transform, and converting the noise-reduced power generation data into standardized power generation data;

[0102] Constructing a hybrid prediction model, wherein the period decomposition unit of the hybrid prediction model uses a variational mode decomposition method to decompose the standardized power generation data into a trend component and a fluctuation component, the trend prediction unit of the hybrid prediction model uses a long short-term memory network structure to model the trend component, and the fluctuation prediction unit of the hybrid prediction model uses a gated recurrent neural network structure to model the fluctuation component;

[0103] The meteorological forecast data is input into the hybrid forecast model, and the hybrid forecast model is optimized and trained based on the combined loss function of the root mean square error and the mean absolute error. The meteorological forecast data corresponding to the time period to be predicted is input into the trained hybrid forecast model to obtain the predicted power generation data.

[0104] During the data preprocessing phase, the system first collects historical renewable energy power generation data, which often contains noise and outliers. To improve prediction accuracy, wavelet transform is used to reduce noise in the power generation data. Specifically, the db4 wavelet basis function is selected to perform a four-level decomposition of the raw power generation data, generating approximate coefficients and detail coefficients. Soft thresholding is then performed on the detail coefficients using a threshold function set at 0.6 times the standard deviation. Wavelet reconstruction is then performed using these processed coefficients to obtain the de-noised power generation data.

[0105] For a wind farm, data from 96 power points collected at 10-minute intervals fluctuated between 0 and 50 MW and contained random noise. Wavelet denoising smoothed the data curve, removing the impact of short-term random fluctuations while preserving the main trends in power generation. After denoising, the data was normalized, mapping the values ​​to the [0, 1] interval. This was done by subtracting the minimum power value from the current power value and dividing it by the difference between the maximum and minimum power values ​​to obtain the normalized power data.

[0106] During the model construction phase, a hybrid forecasting model was designed, consisting of three core components: a period decomposition unit, a trend prediction unit, and a fluctuation prediction unit. The period decomposition unit uses variational mode decomposition to decompose the standardized power data into trend and fluctuation components. During implementation, the standardized power data was decomposed using a mode number K of 3, a penalty factor α of 2000, and a tolerance ε of 1e-7. After decomposition, the first mode function obtained was used as the fluctuation component, and the sum of the remaining mode functions and the residual term was used as the trend component.

[0107] The standardized 96-point power data are input into the variational mode decomposition algorithm, and three mode functions are obtained to reflect the power variation laws of different frequency characteristics. The first mode function contains high-frequency fluctuation information, and the other two mode functions and the residual term reflect the main variation trend of power.

[0108] The trend prediction unit uses a long short-term memory (LSTM) network structure to model trend components. This network consists of an input layer, two LSTM hidden layers, and an output layer. The first LSTM layer has 64 units, and the second layer has 32 units. Each LSTM layer is followed by a dropout layer with a dropout rate of 0.2 to prevent overfitting. Input features include historical trend component data and relevant meteorological forecast data, such as wind speed, wind direction, temperature, and air pressure. A sliding window method is used to construct samples with a window size of 24, meaning that data from the previous 24 time points are used to predict the value at the next time point.

[0109] The volatility prediction unit uses a gated recurrent neural network structure to model the volatility component. This network consists of an input layer, two GRU hidden layers, and an output layer. The first layer has 48 GRU units, and the second layer has 24 GRU units. Each layer is followed by a dropout layer with a dropout rate of 0.2. The input features are the same as those of the trend prediction unit, but the focus is on the ability to capture short-term fluctuations.

[0110] During the model training phase, a combined loss function was used to optimize the hybrid prediction model. This loss function comprehensively considers the root mean square error (RMSE) and mean absolute error (MAE), defining them as a weighted sum with a weight ratio of 0.6:0.4. During training, the Adam optimizer was used, with an initial learning rate of 0.001 and a learning rate decay strategy, reducing the learning rate to 90% of the original rate every 50 epochs. The batch size was set to 32, and the number of training epochs was 200. To prevent overfitting, an early stopping strategy was used, stopping training after 10 consecutive epochs of no improvement in the validation set loss function.

[0111] During the forecast implementation phase, the meteorological forecast data corresponding to the forecasted time period is fed into the trained hybrid forecast model. First, the forecast data undergoes the same normalization process as the training data. It is then fed into the trend prediction unit and the fluctuation prediction unit, respectively, to obtain the trend component and fluctuation component forecast values ​​for the future time period. These two forecast values ​​are summed to obtain the normalized power generation forecast. Finally, the forecast results are denormalized, mapping the predicted values ​​in the [0, 1] interval back to the original power range to obtain the final predicted power generation data.

[0112] Experimental results show that for a 24-hour ahead forecast of a 100MW photovoltaic power plant, this method achieves a mean absolute percentage error of 3.2% and a root mean square error of 2.8MW, improving forecast accuracy by over 15% compared to traditional single-model approaches. The advantage in forecast accuracy is even more pronounced under weather conditions with significant power fluctuations, accurately capturing power surge points and fluctuation trends, providing a reliable basis for grid scheduling and energy market transactions.

[0113] Figure 3 Schematic diagram of the comparison of root mean square errors of different prediction models according to the embodiment of the present invention:

[0114] This figure compares the root mean square error (RMSE) of three different forecasting schemes over eight forecast days. As can be seen, the RMSE of our proposed solution remains consistently low, fluctuating between 2.2% and 2.6% throughout the forecast period, demonstrating the most stable and superior performance. The RMSE of the traditional VMD-LSTM solution is intermediate, fluctuating between 4.5% and 4.9%, nearly double that of our proposed solution. The single LSTM solution performs the worst, with a relatively large RMSE fluctuation between 5.7% and 6.2%. The difference between the three schemes is particularly pronounced on the fourth forecast day: 2.6% for our proposed solution, 4.9% for the traditional VMD-LSTM solution, and 6.0% for the single LSTM solution. The overall trend demonstrates that our proposed solution significantly outperforms the other two in both forecast accuracy and stability, demonstrating its significant advantages in renewable energy power forecasting.

[0115] In an optional embodiment, a hybrid energy storage optimization scheduling model is constructed after feature enhancement is performed based on the predicted electricity load data and the predicted power generation data, and initial charge and discharge power instructions are output based on the hybrid energy storage optimization scheduling model, including:

[0116] constructing a multidimensional spatiotemporal state space based on the predicted electricity load data and the predicted power generation data, and performing feature enhancement on the multidimensional spatiotemporal state space using a spatiotemporal attention mechanism to obtain an enhanced state space;

[0117] Constructing a hierarchical deep reinforcement learning network, wherein the state encoder of the hierarchical deep reinforcement learning network uses a dual-stream graph convolutional neural network to respectively extract the temporal features and spatial features of the enhanced state space; the action generator of the hierarchical deep reinforcement learning network fuses the temporal features with the spatial features based on a gated recurrent unit to generate a charge-discharge power decision; and the value evaluator of the hierarchical deep reinforcement learning network uses a dual evaluation network structure to evaluate the value function of the charge-discharge power decision;

[0118] Based on historical scheduling data, an immediate reward term containing multiple costs is constructed. A recurrent neural network is used to model the charging and discharging operation sequence to obtain a long-term reward term. The immediate reward term is dynamically adjusted through adaptive weight optimization to obtain an adaptive hybrid reward function.

[0119] Combining the adaptive hybrid reward function with the hierarchical deep reinforcement learning network to construct a hybrid energy storage optimization scheduling model, and training the hybrid energy storage optimization scheduling model;

[0120] The current state information is input into the trained hybrid energy storage optimization scheduling model, and the initial charge and discharge power instructions are output through the action generator.

[0121] Obtain forecasted electricity load and power generation data with 24-hour time granularity. This data typically comes from a power grid monitoring system or energy management platform. For example, in an industrial park, the forecasted electricity load peaks at 15MW between 8:00 AM and 6:00 PM on weekdays, while photovoltaic power generation reaches a peak of 10MW between 11:00 AM and 2:00 PM. Wind power generation is relatively high at night and in the early morning, averaging 3MW.

[0122] Based on these prediction data, the system constructs a multi-dimensional spatiotemporal state space. This state space contains a time dimension (24 hours) and a spatial dimension (including the geographical distribution of load nodes, power generation nodes, and energy storage nodes). For example, 10 load nodes, 5 power generation nodes, and 3 hybrid energy storage nodes are arranged in the power grid to form a complex network topology. The system uses a spatiotemporal attention mechanism to enhance the features of this multi-dimensional state space. Specifically, for the time dimension, self-attention is used to calculate the association weights between different moments. For the spatial dimension, the topological relationship between different nodes is captured through a graph attention network, and an enhanced state space representation is finally obtained. In the experiment, this mechanism increased the model's prediction accuracy for peak and valley periods from 85% to 93%.

[0123] In constructing a hierarchical deep reinforcement learning network, the state encoder employs a two-stream graph convolutional neural network architecture. The temporal feature extraction stream uses a 10-layer temporal graph convolutional network with a kernel size of 3 and a hidden layer dimension of 128 to capture 24-hour load and power generation trends. The spatial feature extraction stream employs an 8-layer spatial graph convolutional network with a node embedding dimension of 64 and an edge feature dimension of 32 to capture the interactions between nodes in the power grid topology.

[0124] The action generator, based on a two-layer gated recurrent unit (GRU) network with 128 hidden units, receives fused spatiotemporal features as input and generates charging and discharging power decisions for the hybrid energy storage system. These decisions are made at one-hour intervals, with the output power value for each energy storage device ranging from -5MW to 5MW, where positive values ​​indicate discharge and negative values ​​indicate charge. The value evaluator utilizes a dual evaluation network structure, comprising a main network and a target network. Both networks consist of three fully connected neural networks with 256, 128, and 64 hidden layer nodes, respectively, to assess the long-term value of the current charging and discharging decision. This dual structure effectively reduces the bias in Q-value estimation, making the training process more stable.

[0125] To construct a reasonable reward function, the system designs multiple immediate cost items based on historical dispatch data, including electricity purchase costs (peak electricity price of 0.9 yuan / kWh, off-peak electricity price of 0.4 yuan / kWh), energy storage loss costs (lithium battery cycle efficiency of 90%, lead-acid battery cycle efficiency of 85%), energy storage life costs (deep cycle number corresponds to different loss coefficients), and load fluctuation penalties (a penalty term is added when the load change rate exceeds 10% / hour). The system uses a recurrent neural network with a three-layer LSTM structure and a hidden layer size of 64 to model a continuous 24-hour sequence of charge and discharge operations to obtain a long-term reward term that reflects the quality of the long-term strategy. Through an adaptive weight optimization mechanism, the system dynamically adjusts the weights of each reward item based on the current grid load. For example, the weight of the peak shaving reward is increased during peak load periods, and the weight of the smoothing reward is increased when renewable energy fluctuates significantly. This combination results in an adaptive hybrid reward function.

[0126] By combining an adaptive hybrid reward function with a hierarchical deep reinforcement learning network, a complete hybrid energy storage optimization scheduling model was constructed. The model was trained using an experience replay mechanism with a batch size of 256, a learning rate of 0.001, and a discount factor of 0.95. Training was performed on a training set containing 12 months of historical data, requiring approximately 50,000 iterations to converge. After training, the model achieved a 93% dispatch optimization target achievement rate on the test set, a 15% improvement over traditional methods.

[0127] In practice, the system inputs current state information into the trained hybrid energy storage optimization scheduling model every 15 minutes. This state information includes: current grid load level (e.g., 12MW), renewable energy generation power (e.g., 6MW for photovoltaics and 2MW for wind power), energy storage device status (e.g., lithium battery SOC 65% and lead-acid battery SOC 70%), and electricity price information (currently 0.75 yuan / kWh). After state encoding and decision generation, the model's action generator outputs initial charge and discharge power commands, such as 2.5MW for lithium battery discharge and 1.2MW for lead-acid battery charge, to optimize the overall economic efficiency and stability of the system.

[0128] In an optional embodiment, an immediate reward term containing multiple costs is constructed based on historical scheduling data, a recurrent neural network is used to model the charging and discharging operation sequence to obtain a long-term reward term, and the immediate reward term is dynamically adjusted through adaptive weight optimization to obtain an adaptive hybrid reward function, including:

[0129] Collecting historical dispatch data of the hybrid energy storage system, and constructing an instant reward item including multiple costs based on the historical dispatch data;

[0130] Using a recurrent neural network to perform time series modeling on the charge and discharge operation sequences in the historical scheduling data, extracting long-term dependency features of the charge and discharge operation sequences, and constructing a long-term reward item reflecting the long-term operating benefits of the energy storage system based on the long-term dependency features;

[0131] An evaluation index is constructed based on the system operation state sequence and the cost-benefit sequence. A scheduling effect score of each group of charging and discharging operations in the historical scheduling data is calculated based on the evaluation index. The scheduling effect score is input into an adaptive weight optimizer. The adaptive weight optimizer dynamically adjusts the weight coefficient of each cost sub-item in the instant reward item according to the changing trend of the scheduling effect score. When the scheduling effect score corresponding to a specific cost sub-item decreases, the weight coefficient of the cost sub-item is increased;

[0132] The immediate reward item and the long-term reward item are weightedly combined to obtain an adaptive hybrid reward function.

[0133] like Figure 4 As shown, the method includes:

[0134] The hybrid energy storage system's intelligent dispatch process begins with the collection of historical dispatch data. This data includes, but is not limited to, grid load fluctuations, electricity price fluctuations, the charge and discharge status of energy storage devices, changes in their state of charge (SOC), and the health (SOH) of various types of energy storage devices, as well as corresponding dispatch decision and system revenue data. In practice, the system collects data every five minutes for three consecutive months, totaling approximately 26,000 data samples.

[0135] Based on the collected historical dispatch data, an immediate reward item is constructed. This immediate reward item comprehensively considers multiple cost factors, including: battery cycle degradation cost, electricity purchase cost, peak-valley price difference benefits, demand charge savings, and frequency regulation ancillary service benefits.

[0136] Taking battery cycle degradation costs as an example, a degradation cost model was established by analyzing the relationship between the number of cycles and capacity degradation of energy storage devices at different depths of charge and discharge (DOD). In practical applications, when a lithium battery operates at 80% DOD, its single-cycle degradation cost is approximately 0.02% of the total investment. By converting the depth of charge and discharge of each operation into an equivalent number of full cycles, the degradation cost of a single dispatch can be calculated.

[0137] To capture the long-term dependencies in energy storage system scheduling, this paper employs a recurrent neural network to model historical charging and discharging operation sequences. Specifically, a long short-term memory (LSTM) network architecture is used as input, with a 24-hour continuous charging and discharging operation sequence (one operation point every 5 minutes, for a total of 288 time steps). The network consists of two LSTM layers, each containing 128 hidden units. This architecture captures the temporal patterns and long-term dependencies of the energy storage system's charging and discharging operations.

[0138] The network is trained using 90 consecutive days of historical data, representing approximately 7,776 time series samples. After training, the LSTM network can predict the impact of new operation sequences on the system's future operational performance, thereby constructing reward items that reflect long-term benefits.

[0139] The long-term reward term is constructed based on the LSTM network's predicted output, which includes the predicted cumulative system revenue and device status changes over the next seven days. By comparing the predicted revenue differences under different operation sequences, the contribution of the current operation to long-term benefits is quantified. In actual testing, for peak-valley electricity price arbitrage scenarios, a scheduling strategy using long-term rewards increased average monthly revenue by 12.3% compared to a strategy that only considered immediate revenue, while also reducing battery capacity degradation by 8.7%.

[0140] The core innovation of this invention lies in its adaptive weight optimization mechanism. To achieve dynamic weight adjustment, a feedback mechanism based on scheduling effectiveness evaluation is designed. First, evaluation metrics are defined, including daily average rate of return, battery capacity decay rate, and peak load reduction ratio. These metrics are calculated for each set of charge and discharge operations in historical scheduling data to generate a comprehensive scheduling effectiveness score.

[0141] Taking one experimental scenario as an example, the initial weighting for battery cycle degradation costs was set at 0.3, the weighting for electricity purchase costs was set at 0.4, and the weighting for peak-valley electricity price differentials was set at 0.3. During operation, if the system battery capacity degradation rate exceeded the expected target (for example, the monthly degradation rate exceeded 0.5%), the adaptive optimizer adjusted the weight of the battery degradation cost from 0.3 to 0.45, while appropriately reducing the weights of other cost items, so that the system prioritized battery life preservation.

[0142] The adaptive weight optimizer is implemented using a gradient ascent method. Each day, the system calculates the overall performance score for the current scheduling strategy, analyzes the sensitivity of this score to the weights of each cost item, and determines the direction and magnitude of weight adjustments. To prevent system instability caused by drastic weight adjustments, the weight change step size is capped at no more than 15% of the original weight per adjustment, and the sum of all weights remains constant at 1. In actual application, after 30 days of adaptive adjustments, the system's overall scheduling performance score improved by 17.5%, validating the effectiveness of adaptive weight optimization.

[0143] By weighting the immediate and long-term rewards, a complete adaptive hybrid reward function is constructed. The initial combined weights of the two components are set at 0.6 for the immediate reward and 0.4 for the long-term reward. As the system operates and learns, this ratio is dynamically adjusted through the adaptive mechanism described above. In a comprehensive application scenario, the scheduling strategy guided by the hybrid reward function achieved a 15.8% increase in annualized revenue and a 22.3% increase in battery life compared to the traditional fixed-weight approach.

[0144] Through the adaptive hybrid reward function construction method provided by the present invention, the hybrid energy storage system can take into account both short-term economic benefits and long-term operational benefits, continuously optimize the scheduling strategy in a dynamically changing operating environment, and achieve the unity of the economy, stability and sustainability of the energy storage system.

[0145] In an optional embodiment, based on the operating status data of the hybrid energy storage system and the initial charge and discharge power instructions, a two-layer model predictive control framework is used for optimization, and a corrected charge and discharge power instruction is output through the synergy of distributed optimization and a robust controller, including:

[0146] According to the operating status data of the hybrid energy storage system, the energy storage units are grouped to construct a hierarchical system state space, state feature vectors are extracted based on a variational autoencoder, and spatiotemporal features of the state feature vectors are extracted using a hybrid structure of a gated recurrent neural network and a residual convolutional network to construct a multimodal state prediction model;

[0147] Establish a system dynamic response model, use the output of the multimodal state prediction model to perform state estimation through a particle filter algorithm, perform probability distribution fitting on the prediction results of the system dynamic response model, and construct an adaptive constraint evaluator based on the information entropy criterion;

[0148] Constructing a two-layer model predictive control framework, wherein the long-term optimization layer performs predictive optimization based on the multimodal state prediction model, and the short-term execution layer dynamically corrects the initial charge and discharge power instructions according to the prediction error of the long-term optimization layer to obtain an initial correction result;

[0149] Constructing a multi-time-scale optimization objective function based on the initial correction result, decomposing the multi-time-scale optimization objective function using a distributed alternating direction multiplier method to solve subproblems, and inputting the solution results into a robust controller;

[0150] The initial charge and discharge power instructions are collaboratively optimized with the robust controller through the two-layer model predictive control framework to output a corrected charge and discharge power instruction.

[0151] During the actual operation of a hybrid energy storage system, system status data is collected, including information such as the energy storage unit's state of charge, temperature, health status, and grid frequency fluctuations. The system groups energy storage units by type and characteristics to construct a hierarchical system state space. For example, different types of energy storage units, such as lithium batteries, supercapacitors, and flywheels, are divided into fast-response and high-capacity groups based on their response speed and energy density.

[0152] The state data of each group is subjected to feature extraction through a variational autoencoder. The encoder part of the variational autoencoder contains a three-layer neural network with 128, 64, and 32 hidden layer neurons, respectively. The ReLU activation function is used, and the state feature vector is obtained by training the combined loss function of minimizing the reconstruction error and the KL divergence.

[0153] A hybrid structure of a gated recurrent neural network (GRU) and a residual convolutional network (ResNet) is used to extract spatiotemporal features from the extracted state feature vectors. The GRU network contains two hidden layers, each with 128 neurons, to capture temporal dependencies; the residual convolutional network uses three residual blocks, each of which contains two convolutional layers and a skip connection, with a convolution kernel size of 3×3, to extract spatial features. The output of this hybrid structure is fused through a fully connected layer to construct a multimodal state prediction model. This model can simultaneously consider the historical state evolution and spatial correlation of energy storage units to predict the system state in future time periods. In practical applications, the model has achieved an accuracy of 95.2% in predicting the state of charge of lithium battery energy storage units and 93.8% in predicting the power fluctuation of supercapacitor packs.

[0154] Based on the physical characteristics of the energy storage system, a system dynamic response model, encompassing both electrical and thermal characteristics, was established. This model accounts for factors such as the internal resistance, capacity decay, and temperature effects of the energy storage unit, describing its dynamic response characteristics under different operating conditions. The output of the multimodal state prediction model was applied to state estimation using a particle filter algorithm. The particle filter uses 1000 particle samples and probabilistically represents the system state through importance sampling and resampling.

[0155] The prediction results of the system's dynamic response model are fitted with a probability distribution, using a mixed Gaussian distribution to represent uncertainty. An adaptive constraint evaluator is constructed by calculating information entropy. This evaluator dynamically adjusts system operational constraints based on the uncertainty of the state prediction. The greater the uncertainty, the more conservative the constraints imposed, ensuring the safety of system operation.

[0156] A two-layer model predictive control framework was constructed, consisting of a long-term optimization layer and a short-term execution layer. The long-term optimization layer uses a 15-minute time step and a 24-hour prediction horizon. Long-term optimization is performed based on a multimodal state prediction model. The objective function consists of three components: system operating costs, energy storage lifetime loss costs, and grid frequency regulation benefits. The optimization trajectory is solved using a quadratic programming algorithm.

[0157] The short-term execution layer uses a 10-second time step and a 5-minute prediction horizon. It dynamically corrects the initial charge and discharge power commands based on the real-time system state and the prediction error of the long-term optimization layer. The short-term execution layer uses a rolling-horizon optimization strategy, executing only the control commands for the current period at a time. The optimization problem is updated and re-solved in the next control cycle to obtain the initial correction results.

[0158] Based on the initial correction results, a multi-time-scale optimization objective function is constructed. This objective function takes into account the trade-off between short-term system response speed and long-term optimization objectives, and is solved using the distributed alternating direction multiplier method (AD MM). The ADMM algorithm decomposes the optimization problem into multiple subproblems, each corresponding to a different group of energy storage units or a different time-scale objective. In practical applications, ADMM converges to a better solution after 15 iterations, which is 65% more efficient than centralized optimization. The ADMM solution is input into a robust controller designed based on H∞ control theory and can cope with system parameter uncertainties and external disturbances.

[0159] The initial charge and discharge power commands are collaboratively optimized using a two-layer model predictive control framework and a robust controller. The two-layer model predictive control framework generates the optimal trajectory, while the robust controller handles uncertainty and disturbances. The two work together to output corrected charge and discharge power commands. In practical applications, this method has reduced power fluctuations by 15.3%, improved energy storage utilization by 8.6%, and extended the life of the energy storage system by approximately 12.2%, compared to traditional single-layer model predictive control. This approach also ensures the system's rapid response to grid frequency regulation.

[0160] In an optional embodiment, a two-layer model predictive control framework is constructed, wherein the long-term optimization layer performs predictive optimization based on the multimodal state prediction model, and the short-term execution layer dynamically corrects the initial charge and discharge power instructions based on the prediction error of the long-term optimization layer to obtain an initial correction result, including:

[0161] Constructing a multimodal state space, and mapping the multimodal state space into a long-term prediction state set and a short-term execution state set;

[0162] Based on the long-term prediction state set, a recurrent neural network structure with gated skip connections is used to extract multi-scale state time series features, the multi-scale state time series features are input into a prediction optimizer constructed by a conditional generative adversarial network, and a prediction state sequence is output;

[0163] Establishing a short-term execution layer based on the short-term execution state set, wherein the short-term execution layer uses a bidirectional long short-term memory network to calculate the temporal correlation between the predicted state sequence and the actual state sequence to generate a prediction error feature matrix;

[0164] Constructing an adaptive corrector, wherein the adaptive corrector performs multi-dimensional feature extraction on the prediction error feature matrix to obtain multi-dimensional features, and constructs a multi-head cross attention mechanism based on the multi-dimensional features to generate adaptive correction weights;

[0165] A deep neural network with residual connections is used to perform nonlinear feature fusion on the adaptive correction weight, the multi-scale state time series feature, and the initial charge and discharge power instruction to obtain a nonlinear feature fusion result;

[0166] The nonlinear feature fusion result is input into a probability distribution generation network. The probability distribution generation network samples and generates a set of candidate correction coefficients in the latent variable space based on the distribution characteristics of the prediction error feature matrix, evaluates the set of candidate correction coefficients, selects the optimal correction coefficient based on the Bayesian optimization criterion, and dynamically adjusts the initial charge and discharge power instructions using the optimal correction coefficient to obtain an initial correction result.

[0167] A two-layer model predictive control framework is constructed, in which the long-term optimization layer performs predictive optimization based on a multimodal state prediction model, and the short-term execution layer dynamically corrects the initial charge and discharge power instructions based on the prediction errors of the long-term optimization layer. This framework characterizes the operating state of complex power grid systems by constructing a multimodal state space, mapping this state space into a long-term prediction state set and a short-term execution state set. The long-term state set includes hourly variables such as grid load forecasts, renewable energy generation forecasts, and market electricity price forecasts; the short-term state set includes minute- or second-level variables such as battery state of charge, real-time charge and discharge power, and system frequency.

[0168] Based on a set of long-term prediction states, the system employs a recurrent neural network structure with gated skip connections to extract multi-scale state temporal features. This network structure comprises three layers of recurrent units, each with 64 units, and gated units between adjacent layers to control information flow. This skip connection mechanism enables the fusion of features at different time scales, effectively capturing long-term temporal patterns.

[0169] For the 24-hour power load forecasting task, the system slides a 30-minute window, extracting eight key time series features each time to form a feature vector. These multi-scale state time series features are then input into a prediction optimizer constructed using a conditional generative adversarial network. The optimizer consists of a generator and a discriminator. The generator uses a five-layer fully connected network structure, with 128, 256, 512, 256, and 128 neurons per layer, respectively; the discriminator uses a three-layer convolutional network with a kernel size of 3 and 16, 32, and 64 channels, respectively. Through adversarial training, the system can output a predicted state sequence, such as the forecasted values ​​for grid load and renewable energy generation for the next 24 hours.

[0170] Based on the short-term execution state set, the system establishes a short-term execution layer. This execution layer uses a bidirectional long short-term memory network to calculate the temporal correlation between the predicted state sequence and the actual state sequence. The network consists of two LSTM layers with a hidden layer dimension of 128. It analyzes time series data in both forward and backward directions to generate a prediction error feature matrix. Taking the battery energy management system as an example, the system collects the actual grid state every 5 minutes, compares it with the predicted state, and calculates the relative error rate. This error data is organized into a feature matrix of dimension [24×12,10], where 24×12 represents the number of 5-minute granularity time points in a day, and 10 represents the number of monitored state variables.

[0171] An adaptive corrector is further constructed to perform multi-dimensional feature extraction on the prediction error feature matrix. A three-way parallel feature extraction network is employed: the first uses a one-dimensional convolutional network with a kernel size of 3 to extract local temporal features; the second uses global average pooling to capture the global error distribution; and the third uses a densely connected network to learn correlations between errors. After dimensionality is unified, the three feature extraction results are fed into a multi-headed cross-attention mechanism with four attention heads, each with a dimension of 32. This mechanism generates adaptive correction weights that reflect the importance of different state variables at different time points.

[0172] A deep neural network with residual connections performs nonlinear feature fusion of adaptive correction weights, multi-scale state time series features, and initial charge and discharge power commands. The network consists of five residual blocks, each with two convolutional layers and a skip connection. During the fusion process, the correction weights are applied to the state features and concatenated with the initial power commands. The network then processes the resulting fused feature vector, with a dimension of 256.

[0173] The nonlinear feature fusion results are fed into a probability distribution generation network. This network is based on a variational autoencoder architecture, with both the encoder and decoder utilizing three-layer fully connected networks. The encoder maps the fused features into a 64-dimensional latent variable space and estimates the mean and variance parameters of the latent variables. Based on the statistical properties of the prediction error feature matrix, such as a mean of -0.023 and a standard deviation of 0.158, the system samples the latent variable space 100 times to generate a set of candidate correction coefficients.

[0174] For each candidate correction coefficient, the system evaluates its performance through simulations, calculating metrics such as energy efficiency and peak-valley difference smoothing. Based on the Bayesian optimization criterion, the system selects the correction coefficient with the highest overall score from the candidate set as the optimal correction coefficient. For example, in one actual application, the system selected an optimal correction coefficient of 0.87, indicating a 13% reduction in the initial charge and discharge power commands. The system then dynamically adjusts the initial charge and discharge power commands using this optimal correction coefficient to obtain the initial correction result.

[0175] This technical solution, through multimodal information fusion and deep learning methods, enables accurate prediction of grid status and precise control of battery charging and discharging strategies, significantly improving the adaptability and stability of the energy management system. Practical application results show that compared with traditional single-layer control methods, this method improves peak-to-valley difference smoothing efficiency by approximately 18.5% and extends battery life by approximately 12.3%.

[0176] Figure 5 This is a bar chart showing the performance comparison and analysis of the two-layer model predictive control framework according to an embodiment of the present invention:

[0177] The figure compares three different control models (a traditional predictive control model, a single-layer predictive control model, and a two-layer predictive control framework) across five key performance metrics. In terms of prediction accuracy, the two-layer predictive control framework achieved 94.2%, significantly outperforming the single-layer model's 88.7% and the traditional model's 83.5%. In terms of response speed, the two-layer framework also performed best, achieving 92.8%, compared to the single-layer model's 85.4% and the traditional model's 78.2%. In the anti-interference capability assessment, the two-layer framework achieved 89.5%, compared to the single-layer model's 79.3%, with the traditional model having the lowest score of 71.6%. In terms of energy utilization, the two-layer framework achieved 93.1%, far exceeding the single-layer model's 84.6% and the traditional model's 80.3%. In terms of regulation stability, the two-layer framework maintained its lead, achieving 91.7%, compared to the single-layer model's 82.1% and the traditional model's 75.8%. The data demonstrates that the two-layer predictive control framework exhibits significant performance advantages across all evaluation metrics, offering superior overall performance compared to the other two models.

[0178] A second aspect of an embodiment of the present invention provides a data center hybrid energy storage and renewable energy intelligent dispatching and control system, including:

[0179] The first unit is used to collect data on the power load of the data center, the operating status data of the hybrid energy storage system, and the power generation data of renewable energy;

[0180] The second unit is configured to extract time series features of the power load data using a two-layer temporal memory network, process the time series features using a multi-head attention mechanism to establish a data center power load forecasting model, and predict the data center power load in a future time period based on the power load forecasting model to obtain predicted power load data;

[0181] A third unit is configured to establish a renewable energy power generation prediction model based on the renewable energy power generation data, and predict the renewable energy power generation in a future time period according to the renewable energy power generation prediction model to obtain predicted power generation data;

[0182] A fourth unit is configured to construct a hybrid energy storage optimization scheduling model after feature enhancement based on the predicted electricity load data and the predicted power generation data, and output an initial charge and discharge power instruction based on the hybrid energy storage optimization scheduling model;

[0183] a fifth unit, configured to optimize the hybrid energy storage system based on the operating status data and the initial charge and discharge power instructions using a two-layer model predictive control framework, and output a corrected charge and discharge power instruction through the synergy of distributed optimization and a robust controller;

[0184] The sixth unit is configured to control the charge and discharge operations of the hybrid energy storage system according to the corrected charge and discharge power instructions.

[0185] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0186] processor;

[0187] a memory for storing processor-executable instructions;

[0188] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0189] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0190] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0191] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data center hybrid energy storage and renewable energy intelligent scheduling and control method, characterized in that: include: Collect data center power load data, hybrid energy storage system operating status data, and renewable energy power generation data; A two-layer temporal memory network is used to extract the temporal features of the power load data, and the temporal features are processed through a multi-head attention mechanism to establish a data center power load forecasting model. The power load of the data center in a future time period is predicted based on the power load forecasting model to obtain predicted power load data; Establishing a renewable energy power generation prediction model based on the renewable energy power generation data, and predicting the renewable energy power generation in a future time period according to the renewable energy power generation prediction model to obtain predicted power generation data; Constructing a hybrid energy storage optimization scheduling model after feature enhancement based on the predicted electricity load data and the predicted power generation data, and outputting initial charge and discharge power instructions based on the hybrid energy storage optimization scheduling model; Based on the operating status data of the hybrid energy storage system and the initial charge and discharge power instructions, a two-layer model predictive control framework is used for optimization, and a corrected charge and discharge power instruction is output through the synergy of distributed optimization and a robust controller; The charge and discharge operations of the hybrid energy storage system are controlled according to the corrected charge and discharge power instructions.

2. The method according to claim 1, characterized in that A two-layer temporal memory network is used to extract the temporal features of the power load data. The temporal features are processed through a multi-head attention mechanism to establish a data center power load forecasting model. The power load of the data center in the future time period is predicted based on the power load forecasting model to obtain predicted power load data, including: Preprocessing the power load data, and converting the power load data into standardized load data using a maximum and minimum value normalization method; A double-layer temporal memory network is used to extract the temporal features of the standardized load data, a random dropout layer is set between the two layers of the double-layer temporal memory network, and the temporal features are fused with the time features, the temperature features, and the load features to obtain fused features; Processing the fused features through a multi-head attention mechanism to calculate a multi-head attention weight, where the multi-head attention weight is determined by the attention scores calculated by the multiple attention heads, and the attention scores are calculated by the query vector, the key vector, and the value vector; Performing a weighted operation on the multi-head attention weights and the fusion features to obtain an attention feature, and establishing a data center power load prediction model based on the attention feature; Constructing a loss function including a weighted combination of a root mean square error term and a mean absolute percentage error term, and optimizing and training the data center power load forecasting model based on the loss function; Inputting the characteristic data corresponding to the time period to be predicted into the trained data center power load prediction model to obtain an initial prediction result; The confidence interval of the initial prediction result is calculated based on the historical prediction error distribution, and when the initial prediction result exceeds the confidence interval, correction is performed to obtain predicted electricity load data.

3. The method according to claim 1, characterized in that Establishing a renewable energy power generation prediction model based on the renewable energy power generation data, and predicting the renewable energy power generation in a future time period according to the renewable energy power generation prediction model to obtain predicted power generation data, including: Preprocessing the power generation data of the renewable energy, performing noise reduction processing on the power generation data through wavelet transform, and converting the noise-reduced power generation data into standardized power generation data; Constructing a hybrid prediction model, wherein the period decomposition unit of the hybrid prediction model uses a variational mode decomposition method to decompose the standardized power generation data into a trend component and a fluctuation component, the trend prediction unit of the hybrid prediction model uses a long short-term memory network structure to model the trend component, and the fluctuation prediction unit of the hybrid prediction model uses a gated recurrent neural network structure to model the fluctuation component; The meteorological forecast data is input into the hybrid forecast model, and the hybrid forecast model is optimized and trained based on the combined loss function of the root mean square error and the mean absolute error. The meteorological forecast data corresponding to the time period to be predicted is input into the trained hybrid forecast model to obtain the predicted power generation data.

4. The method according to claim 1, wherein After feature enhancement is performed on the predicted power load data and the predicted power generation data, a hybrid energy storage optimization scheduling model is constructed, and initial charge and discharge power instructions are output according to the hybrid energy storage optimization scheduling model, including: constructing a multidimensional spatiotemporal state space based on the predicted electricity load data and the predicted power generation data, and performing feature enhancement on the multidimensional spatiotemporal state space using a spatiotemporal attention mechanism to obtain an enhanced state space; Constructing a hierarchical deep reinforcement learning network, wherein the state encoder of the hierarchical deep reinforcement learning network uses a dual-stream graph convolutional neural network to respectively extract the temporal features and spatial features of the enhanced state space; the action generator of the hierarchical deep reinforcement learning network fuses the temporal features with the spatial features based on a gated recurrent unit to generate a charge-discharge power decision; and the value evaluator of the hierarchical deep reinforcement learning network uses a dual evaluation network structure to evaluate the value function of the charge-discharge power decision; Based on historical scheduling data, an immediate reward term containing multiple costs is constructed. A recurrent neural network is used to model the charging and discharging operation sequence to obtain a long-term reward term. The immediate reward term is dynamically adjusted through adaptive weight optimization to obtain an adaptive hybrid reward function. Combining the adaptive hybrid reward function with the hierarchical deep reinforcement learning network to construct a hybrid energy storage optimization scheduling model, and training the hybrid energy storage optimization scheduling model; The current state information is input into the trained hybrid energy storage optimization scheduling model, and the initial charge and discharge power instructions are output through the action generator.

5. The method according to claim 4, characterized in that Based on historical scheduling data, an immediate reward term containing multiple costs is constructed. A recurrent neural network is used to model the charging and discharging operation sequence to obtain a long-term reward term. The immediate reward term is dynamically adjusted through adaptive weight optimization. The combination results in an adaptive hybrid reward function, including: Collecting historical dispatch data of the hybrid energy storage system, and constructing an instant reward item including multiple costs based on the historical dispatch data; Using a recurrent neural network to perform time series modeling on the charge and discharge operation sequences in the historical scheduling data, extracting long-term dependency features of the charge and discharge operation sequences, and constructing a long-term reward item reflecting the long-term operating benefits of the energy storage system based on the long-term dependency features; An evaluation index is constructed based on the system operation state sequence and the cost-benefit sequence. A scheduling effect score of each group of charging and discharging operations in the historical scheduling data is calculated based on the evaluation index. The scheduling effect score is input into an adaptive weight optimizer. The adaptive weight optimizer dynamically adjusts the weight coefficient of each cost sub-item in the instant reward item according to the changing trend of the scheduling effect score. When the scheduling effect score corresponding to a specific cost sub-item decreases, the weight coefficient of the cost sub-item is increased; The immediate reward item and the long-term reward item are weightedly combined to obtain an adaptive hybrid reward function.

6. The method according to claim 1, characterized in that Based on the operating status data of the hybrid energy storage system and the initial charge and discharge power instructions, a two-layer model predictive control framework is used for optimization, and a corrected charge and discharge power instruction is output through the synergy of distributed optimization and a robust controller, including: According to the operating status data of the hybrid energy storage system, the energy storage units are grouped to construct a hierarchical system state space, state feature vectors are extracted based on a variational autoencoder, and spatiotemporal features of the state feature vectors are extracted using a hybrid structure of a gated recurrent neural network and a residual convolutional network to construct a multimodal state prediction model; Establish a system dynamic response model, use the output of the multimodal state prediction model to perform state estimation through a particle filter algorithm, perform probability distribution fitting on the prediction results of the system dynamic response model, and construct an adaptive constraint evaluator based on the information entropy criterion; Constructing a two-layer model predictive control framework, wherein the long-term optimization layer performs predictive optimization based on the multimodal state prediction model, and the short-term execution layer dynamically corrects the initial charge and discharge power instructions according to the prediction error of the long-term optimization layer to obtain an initial correction result; Constructing a multi-time-scale optimization objective function based on the initial correction result, decomposing the multi-time-scale optimization objective function using a distributed alternating direction multiplier method to solve subproblems, and inputting the solution results into a robust controller; The initial charge and discharge power instructions are collaboratively optimized with the robust controller through the two-layer model predictive control framework to output a corrected charge and discharge power instruction.

7. The method according to claim 6, characterized in that A two-layer model predictive control framework is constructed, wherein the long-term optimization layer performs predictive optimization based on the multimodal state prediction model, and the short-term execution layer dynamically corrects the initial charge and discharge power instructions according to the prediction error of the long-term optimization layer to obtain an initial correction result, including: Constructing a multimodal state space, and mapping the multimodal state space into a long-term prediction state set and a short-term execution state set; Based on the long-term prediction state set, a recurrent neural network structure with gated skip connections is used to extract multi-scale state time series features, the multi-scale state time series features are input into a prediction optimizer constructed by a conditional generative adversarial network, and a prediction state sequence is output; Establishing a short-term execution layer based on the short-term execution state set, wherein the short-term execution layer uses a bidirectional long short-term memory network to calculate the temporal correlation between the predicted state sequence and the actual state sequence to generate a prediction error feature matrix; Constructing an adaptive corrector, wherein the adaptive corrector performs multi-dimensional feature extraction on the prediction error feature matrix to obtain multi-dimensional features, and constructs a multi-head cross attention mechanism based on the multi-dimensional features to generate adaptive correction weights; A deep neural network with residual connections is used to perform nonlinear feature fusion on the adaptive correction weight, the multi-scale state time series feature, and the initial charge and discharge power instruction to obtain a nonlinear feature fusion result; The nonlinear feature fusion result is input into a probability distribution generation network. The probability distribution generation network samples and generates a set of candidate correction coefficients in the latent variable space based on the distribution characteristics of the prediction error feature matrix, evaluates the set of candidate correction coefficients, selects the optimal correction coefficient based on the Bayesian optimization criterion, and dynamically adjusts the initial charge and discharge power instructions using the optimal correction coefficient to obtain an initial correction result.

8. A data center hybrid energy storage and renewable energy intelligent dispatching and control system, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to collect data on the power load of the data center, the operating status data of the hybrid energy storage system, and the power generation data of renewable energy; The second unit is configured to extract time series features of the power load data using a two-layer temporal memory network, process the time series features using a multi-head attention mechanism to establish a data center power load forecasting model, and predict the data center power load in a future time period based on the power load forecasting model to obtain predicted power load data; A third unit is configured to establish a renewable energy power generation prediction model based on the renewable energy power generation data, and predict the renewable energy power generation in a future time period according to the renewable energy power generation prediction model to obtain predicted power generation data; A fourth unit is configured to construct a hybrid energy storage optimization scheduling model after feature enhancement based on the predicted electricity load data and the predicted power generation data, and output an initial charge and discharge power instruction based on the hybrid energy storage optimization scheduling model; a fifth unit, configured to optimize the hybrid energy storage system based on the operating status data and the initial charge and discharge power instructions using a two-layer model predictive control framework, and output a corrected charge and discharge power instruction through the synergy of distributed optimization and a robust controller; The sixth unit is configured to control the charge and discharge operations of the hybrid energy storage system according to the corrected charge and discharge power instructions.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images