Energy storage power station anti-countercurrent control method based on combined model

Abstract

The invention provides an energy storage power station anti-countercurrent control method based on a combined model. The method comprises the following steps: S1, data acquisition and preprocessing; s2, constructing a bidirectional long short-term memory network-random forest prediction model; s3, a model training and prediction process; and S4, evaluating and adjusting the model. By constructing the TSSA-bidirectional long-short-term memory network-random forest prediction model and a related optimization algorithm, the power prediction precision can be improved by integrating multiple factors, the power change trend and the nonlinear relationship can be accurately captured, local optimum is avoided, and a reliable basis is provided for an anti-countercurrent control decision; therefore, real-time accurate decision making is assisted, complex working conditions are effectively handled, the output power of the energy storage converter is adjusted in advance, countercurrent is prevented, and the stability of a power grid is guaranteed. Finally, the overall operation efficiency and safety of the energy storage power station are improved, the charging and discharging strategy is optimized, the energy loss is reduced, the service life of equipment is prolonged, the maintenance cost is reduced, and the operation reliability and economy of the energy storage power station in a power system are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of energy storage power station control, and in particular to an energy storage power station backflow prevention control method based on a combined model. Background Art

[0002] With the widespread use of renewable energy (such as solar energy, wind energy, etc.) in power systems, its intermittency and volatility have brought challenges to the stable operation of the power grid. As an effective solution, energy storage power stations can store excess electricity and release it when needed, playing the role of "peak shaving and valley filling", improving the power quality and stability of the power grid.

[0003] In the interaction between energy storage power stations and power grids, reverse flow may have adverse effects on the power grid, such as voltage fluctuations, increased harmonics, etc., and may also affect the safety and life of the energy storage power station itself. Therefore, it is crucial to accurately predict the power changes of the energy storage system and achieve effective anti-reverse flow control. Traditional power prediction methods, such as time series analysis and regression analysis, often have difficulty accurately capturing the nonlinear characteristics and time series dependencies in the data when processing complex operating data of energy storage power stations. The power of energy storage power stations is affected by many factors, including ambient temperature, battery status, grid load, etc. There are complex nonlinear relationships between these factors. Traditional methods are difficult to fully consider the combined effects of these factors, resulting in limited prediction accuracy. A single prediction model has poor adaptability when facing the operation of energy storage power stations under different working conditions. For example, when the grid load suddenly changes or the environmental conditions change dramatically, the model may not be able to adjust the prediction results in a timely and accurate manner, thus affecting the effect of anti-reverse flow control.

[0004] Some existing model optimization algorithms tend to fall into local optimality when searching for the optimal solution, and are unable to find the most suitable model parameter combination for energy storage power station power prediction in the global scope. This limits the performance improvement of the model and makes it difficult to meet the needs of energy storage power stations for high-precision power prediction and stable anti-backflow control. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a backflow prevention control method for an energy storage power station based on a combined model.

[0006] To achieve the above purpose, the present invention adopts the following technical solution: a method for preventing backflow in an energy storage power station based on a combined model, wherein the specific steps of the method are as follows:

[0007] S1: Data collection and preprocessing;

[0008] S11: Grid power data collection:

[0009] The grid power detection device collects grid power data in real time, including three-phase grid power and its direction information; at the same time, the energy storage converter obtains its actual output power and the operating parameters of the energy storage system, including battery voltage and current;

[0010] S12: Environmental and operational data collection:

[0011] Collect environmental data such as ambient temperature and humidity, as well as the operating status information of the energy storage power station, including charging and discharging mode and remaining battery capacity;

[0012] S13: Outlier processing:

[0013] Using a method based on statistical principles to detect and process outliers on the data collected in step S11 and step S12, and calculating the mean and standard deviation;

[0014] S14: Data filtering and smoothing:

[0015] The power data is filtered using a sliding average filtering algorithm.

[0016] S2: Construct a bidirectional long short-term memory network-random forest prediction model;

[0017] It includes building a bidirectional long short-term memory network, random forest prediction and TSSA optimization algorithm; the specific steps of building a bidirectional long short-term memory network-random forest prediction model are as follows:

[0018] S21: Building a bidirectional long short-term memory network:

[0019] The bidirectional long short-term memory network is constructed by two LSTM layers, the forward LSTM layer and the backward LSTM layer. The bidirectional long short-term memory network extracts features that are important for power prediction based on factors such as past power change trends and possible future environmental impacts.

[0020] S22: Construct random forest prediction:

[0021] Since the operation data of energy storage power stations is complex, affected by multiple factors and has nonlinear relationships, multiple decision trees of random forests can learn and make decisions on different feature combinations. And since random forests are composed of multiple decision trees, the power of energy storage power stations is predicted based on the features extracted by the bidirectional long short-term memory network.

[0022] S23: Construct TSSA optimization algorithm;

[0023] The specific steps to build the TSSA optimization algorithm are as follows:

[0024] S231: Population initialization:

[0025] Circle chaos mapping is used to initialize the sparrow population, improve the edge aggregation of sparrows, and improve the quality of the initial solution to better explore the feature space of energy storage power station related data;

[0026] S232: Location Update:

[0027] The location update includes the discoverer location update, the joiner location update, and the alerter location update;

[0028] The discoverer position update is usually responsible for detecting, identifying or capturing certain key events. When the target location or related information changes, the discoverer will update the target location known to it;

[0029] The joiner position update refers to following or joining other sparrows after discovering the target, usually entering the task execution area under the guidance of the discoverer; the joiner updates its own geographical location or status information according to the current task requirements or changes in the target location;

[0030] The position update of the early warning device refers to the sparrows with early warning performance, which account for 10% to 20% of the sparrow population. The initial position is randomly generated by the system, and when danger is detected, its position is updated;

[0031] S233: Add disturbance:

[0032] The added disturbance refers to calculating the fitness of the population after the position is updated, finding the sparrow with the highest fitness value, and adopting the firefly disturbance strategy to update the optimal solution.

[0033] S3: training and predicting the prediction model;

[0034] The TSSA algorithm is used to optimize the hyperparameters of the bidirectional LSTM-random forest model to minimize the prediction error.

[0035] The specific steps of training and predicting the prediction model are as follows:

[0036] S31: Hyperparameter Optimization:

[0037] The TSSA algorithm is used to optimize the hyperparameters in the bidirectional long short-term memory network-random forest model, including the number of hidden layers in the bidirectional long short-term memory network layer, the learning rate, and the number of decision trees in the random forest; the optimization goal is to minimize the prediction error;

[0038] S32: Model training process:

[0039] The preprocessed training set data is input into the bidirectional long short-term memory network, and forward and backward propagation is performed according to the calculation formula of the above bidirectional long short-term memory network to calculate the hidden layer state and output features; the output features of the bidirectional long short-term memory network are then used as the input of the random forest model, and samples and features are randomly extracted to construct a decision tree according to the training mechanism of the random forest; in each iteration, the prediction error of the model on the training set is recalculated according to the TSSA algorithm update, that is, the hyperparameter combination, and the fitness value of the hyperparameter group is adjusted according to the error; S33: real-time data input and processing:

[0040] The pre-processed real-time data, including the current power, environmental data and operating status data, is input into the trained TSSA bidirectional long short-term memory network and random forest model; the real-time data also needs to be normalized and pre-processed in the same way as the training data S34: Feature extraction and prediction:

[0041] The bidirectional long short-term memory network layer in the model first extracts features from the input real-time data, and calculates the hidden layer state and output feature vector according to the bidirectional long short-term memory network calculation formula mentioned above; then the extracted feature vector is input into the random forest model, and the random forest model makes predictions based on the trained decision tree.

[0042] S4: Evaluate and adjust the prediction model.

[0043] The specific steps of evaluating and adjusting the prediction model are as follows:

[0044] S41: Evaluation index calculation:

[0045] During model training and prediction, the test set data is used regularly to evaluate the model; S42: Model adjustment strategy:

[0046] According to the results of the evaluation indicators in step S41, if it is found that the model performance has decreased or the prediction error has increased, the model needs to be adjusted.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] The present invention discloses a method for preventing backflow control of an energy storage power station based on a combined model, which has the following beneficial effects: by constructing a TSSA-BiLSTM-RF prediction model and related optimization algorithms, it can comprehensively improve the accuracy of power prediction, accurately capture power change trends and nonlinear relationships, avoid falling into local optimality, and provide a reliable basis for preventing backflow control decisions; thereby facilitating real-time and accurate decision-making, effectively coping with complex working conditions, adjusting the output power of the energy storage converter in advance, preventing backflow, and ensuring the stability of the power grid; ultimately achieving the improvement of the overall operating efficiency and safety of the energy storage power station, optimizing the charging and discharging strategies, reducing energy loss, extending the service life of the equipment, reducing maintenance costs, and enhancing its operating reliability and economy in the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A schematic diagram of a method framework in one embodiment of the present invention;

[0050] Figure 2 Schematic diagram of the model training process framework in one embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to provide a further understanding of the purpose, structure, features, and functions of the present invention, the following detailed description is given in conjunction with the embodiments.

[0052] Please refer to Figure 1 as well as Figure 2 The present invention provides a method for preventing backflow in an energy storage power station based on a combined model, and the specific steps of the method are as follows:

[0053] S1: Data collection and preprocessing;

[0054] S11: Grid power data collection:

[0055] The grid power detection device collects grid power data in real time, including three-phase grid power and its direction information; at the same time, the energy storage converter obtains its actual output power and the operating parameters of the energy storage system, including battery voltage and current.

[0056] S12: Environmental and operational data collection:

[0057] Collect environmental data such as ambient temperature and humidity, as well as operating status information of the energy storage power station, including charging and discharging modes and remaining battery capacity.

[0058] S13: Outlier processing:

[0059] Using a method based on statistical principles to detect and process outliers on the data collected in step S11 and step S12, and calculating the mean μ and standard deviation σ;

[0060] If a data point exceeds the range of [μ-kσ,μ+kσ] (k is a coefficient set according to actual conditions, generally around 3), the data point is considered an outlier. For outliers, the average value of the adjacent data points is used to replace them.

[0061] S14: Data filtering and smoothing:

[0062] Use sliding average filtering algorithm to filter the power data;

[0063] Assume the filter window size is n, the power data after filtering P filtered The calculation formula of (t) is: Where P(i) is the original power data sequence. S2: Construct a bidirectional long short-term memory network-random forest prediction model (i.e., TSSA-BiLSTM-RF model);

[0064] It includes building a bidirectional long short-term memory network, random forest prediction and TSSA optimization algorithm; the specific steps of building a bidirectional long short-term memory network-random forest prediction model are as follows:

[0065] S21: Building a bidirectional long short-term memory network (i.e. BiLSTM):

[0066] The bidirectional long short-term memory network is constructed by two LSTM layers, the forward LSTM layer and the backward LSTM layer. The bidirectional long short-term memory network extracts features that are important for power prediction based on factors such as past power change trends and possible future environmental impacts.

[0067] The hidden state of the forward LSTM layer at time t Update formula:

[0068]

[0069] Where f is the activation function (such as tanh function), is the input data x t to the weight matrix of the forward hidden state;

[0070] is the forward hidden state at the previous moment The weight matrix of the forward hidden state to the current moment; is the bias vector.

[0071] Hidden state of the backward LSTM layer The calculation formula is:

[0072]

[0073] here and They are the weight matrix and bias vector corresponding to the backward LSTM layer respectively.

[0074] The final output of the BiLSTM layer at time t

[0075] S22: Construct random forest prediction (i.e. RF prediction):

[0076] Since the operating data of energy storage power stations is complex, affected by multiple factors and has nonlinear relationships, the multiple decision trees of random forests can learn and make decisions on different feature combinations. And since random forests are composed of multiple decision trees, the power of energy storage power stations is predicted based on the features extracted by the bidirectional long short-term memory network.

[0077] For regression problems, suppose the prediction value of the i-th decision tree is y i , the number of decision trees is n est , then the predicted output of RF is y RF The calculation formula is During the training process, each decision tree is built based on a random subset of the training data, samples are randomly drawn with replacement from the original training data set, and some features are randomly selected to train the decision tree.

[0078] S23: Construct TSSA optimization algorithm;

[0079] The specific steps to build the TSSA optimization algorithm are as follows:

[0080] S231: Population initialization:

[0081] Circle chaos mapping is used to initialize the sparrow population, improve the edge aggregation of sparrows, and improve the quality of the initial solution to better explore the feature space of energy storage power station related data;

[0082] The Circle chaos mapping formula is:

[0083]

[0084] Where i is the dimension of the solution, λ 1 ,λ 2 ,λ 3 is a constant.

[0085] S232: Location Update:

[0086] The location update includes the discoverer location update, the joiner location update, and the alerter location update;

[0087] The discoverer position update is usually responsible for detecting, identifying or capturing certain key events. When the target position or related information changes, the discoverer will update the target position known to it.

[0088] Finder position update: The finder with the highest fitness value updates its position according to the following formula:

[0089]

[0090] Among them, Q represents a random number in the range of 0-1 that follows a normal distribution, t represents the number of iterations, α∈(0,1) is a random number, L represents a matrix of all 1s, and R 2 is the warning value, iter max represents the maximum number of iterations, ST represents the safety threshold, and X i,j Represents the value of the i-th sparrow in the j-th dimension.

[0091] The joiner position update refers to following or joining other sparrows after discovering the target, usually entering the task execution area under the guidance of the discoverer; the joiner updates its own geographical location or status information according to the current task requirements or changes in the target location;

[0092] Joiner position update: The joiner updates the position according to the sine and cosine algorithm. The formula is as follows:

[0093]

[0094] in, l 2 Any random number that follows uniform distribution in ∈(0,2π), l 3 ∈(0,2) any random number that follows a uniform distribution, l 4 ∈(0,1) is any random number that follows a uniform distribution, t is the current iteration number, xb d represents the sparrow is in the best position in the dth dimension, iter max is the maximum number of iterations.

[0095] The position update of the early warning device refers to the sparrows with early warning performance, which account for 10% to 20% of the sparrow population. The initial position is randomly generated by the system, and when danger is detected, its position is updated;

[0096] Warning signal position update: The position update formula is:

[0097]

[0098] Among them, xb d represents the best position of the sparrow in the dth dimension, K∈[-1,1] is a random number, e represents a very small number, β represents the step size control parameter, and f i Represents the fitness value of the current sparrow, f g ,f w Represent the current highest and lowest values ​​of global energy respectively.

[0099] S233: Add disturbance:

[0100] The added disturbance refers to calculating the fitness of the population after the position is updated, finding the sparrow with the highest fitness value, and adopting the firefly disturbance strategy to update the optimal solution.

[0101] Calculate the fitness of the population after the position update, find the sparrow with the highest fitness value, and the light sensitivity of the firefly is

[0102] Among them, L 0 is the initial light brightness, the higher the energy, the greater the discovery labor; γ is the light intensity absorption coefficient, the greater the distance between individual sparrows, the greater the γ will gradually weaken; r ij Represents the spatial distance between sparrows i and j;

[0103] The attraction between individual sparrows

[0104] Among them, β 0 Indicates the initial attractiveness of the sparrow;

[0105] The firefly position update formula is:

[0106] S′ i =S i +β(r)(S i -S j )+αrand();

[0107] Among them, S i and S j , represents the current spatial position of sparrow i and sparrow j; α represents the perturbation step range is 0-1; rand∈(-0.5,0.5) is a uniformly distributed random number.

[0108] During the operation of energy storage power stations, complex operating conditions such as frequent fluctuations in the power grid and sudden changes in the environment may cause traditional models to fall into local optimality and fail to accurately predict power.

[0109] By introducing the firefly perturbation strategy, when the model tends to fall into the local optimal solution when dealing with the complex operating conditions of the energy storage power station, it can use the perturbation to jump out of the local optimal solution and continue to search for a better solution in the global scope, thereby improving the model's ability to predict power changes under different operating conditions of the energy storage power station.

[0110] Assume that in a certain iteration, it is found that the fitness value of the current optimal sparrow (corresponding to a set of hyperparameter combinations) is high but no longer significantly improved. At this time, the light sensitivity brightness L(r) and attraction β(r) between each sparrow in the population are calculated.

[0111] For light brightness, if the two sparrows are close to each other (r ij When the distance is small), the light intensity absorption coefficient γ is relatively small and L(r) is large, indicating that they have a greater influence on each other; conversely, when the distance is far, L(r) is small.

[0112] The attraction β(r) is also related to the distance, the closer the distance, the greater the attraction.

[0113] According to S′ i =S i +β(r)(S i -S j )+αrand(),

[0114] Update the position of the optimal sparrow, β(r)(S i -S j ) makes the optimal sparrow be attracted or repelled by other sparrows, and its direction and magnitude depend on the attraction β(r) and the difference in position with other sparrows (S i -S j ), and αrand() introduces random disturbances, so that the optimal sparrow has the opportunity to jump out of the local optimal area.

[0115] For example, if the optimal sparrow is currently near a locally optimal hyperparameter combination, it may move to a new position through interactions with surrounding sparrows and random perturbations. The hyperparameter combination corresponding to this position may bring better prediction results, thereby continuing to search for a better solution globally.

[0116] S3: training and predicting the prediction model;

[0117] The TSSA algorithm is used to optimize the hyperparameters of the bidirectional LSTM-random forest model to minimize the prediction error.

[0118] The specific steps of training and predicting the prediction model are as follows:

[0119] S31: Hyperparameter Optimization:

[0120] Use the TSSA algorithm to optimize the hyperparameters in the bidirectional long short-term memory network-random forest model, including the number of hidden layers n in the bidirectional long short-term memory network layer h , learning rate η, number of decision trees in random forest n est ; The optimization goal is to minimize the prediction error.

[0121] The root mean square error (RMSE) is used as the error evaluation index, and the RMSE calculation formula is:

[0122]

[0123] Where m is the number of samples, y i is the true value, is the predicted value. At the same time, the mean absolute error (MAE) is introduced as an auxiliary evaluation indicator. The MAE calculation formula is: By comprehensively considering RMSE and MAE, the prediction performance of the model can be more comprehensively evaluated. In addition, the coefficient of determination R 2 To measure the goodness of fit of the model, R 2 The calculation formula is

[0124] Where y is the mean of the true values.

[0125] During the training process, the model parameters are continuously adjusted until the stopping condition is met (such as reaching the maximum number of iterations T max Or error convergence, that is, the error change after k consecutive iterations is less than the set threshold ε). In order to avoid overfitting, the early stopping method can be used, that is, the training is stopped when the error on the validation set no longer decreases.

[0126] S32: Model training process:

[0127] The preprocessed training set data is input into the bidirectional long short-term memory network, and forward and backward propagation is performed according to the calculation formula of the above bidirectional long short-term memory network to calculate the hidden layer state and output features; the output features of the bidirectional long short-term memory network are then used as the input of the random forest model, and samples and features are randomly extracted to construct a decision tree according to the training mechanism of the random forest; in each iteration, the prediction error of the model on the training set is recalculated according to the TSSA algorithm update, that is, the hyperparameter combination, and the fitness value of the hyperparameter group is adjusted according to the error.

[0128] For example, if the prediction error corresponding to a certain hyperparameter combination is small, the fitness value of the sparrow corresponding to this combination is high, and vice versa. Through continuous iteration, the TSSA algorithm guides the sparrow to search in the hyperparameter space and gradually find the hyperparameter combination that minimizes the prediction error. In this process, the best hyperparameter combination and its corresponding error index of each iteration are recorded for subsequent analysis and comparison.

[0129] S33: Real-time data input and processing:

[0130] The preprocessed real-time data, including the current power, environmental data, and operating status data, is input into the trained TSSA bidirectional long short-term memory network and random forest model. The real-time data also needs to undergo the same normalization and preprocessing operations as the training data to ensure that the data format and range are consistent with those during model training.

[0131] For example, for real-time power data, if translation and normalization processing are performed during training, it is also processed according to the same translation amount and normalization formula.

[0132] S34: Feature extraction and prediction:

[0133] The bidirectional long short-term memory network layer in the model first extracts features from the input real-time data, and calculates the hidden layer state and output feature vector according to the bidirectional long short-term memory network calculation formula mentioned above; then the extracted feature vector is input into the random forest model, and the random forest model makes predictions based on the trained decision tree.

[0134] Assume that the feature vector of real-time data after BiLSTM processing is F, and the prediction output of the j-th decision tree in the RF model for this feature vector is y j (F), then the final prediction output of the RF model is Get the power prediction value P of the energy storage system at the future time predict .

[0135] In order to improve the stability and accuracy of the forecast, the rolling forecast method can be adopted, that is, the latest actual data is used to update the input sequence each time the forecast is made, and the forecast is made step by step.

[0136] For example, when predicting the power at the next moment, the actual power value at the current moment is added to the input sequence, the oldest data point is removed, and then feature extraction and prediction are performed again. This allows the model to better adapt to the dynamic changes in the operating status of the energy storage power station.

[0137] At the same time, during the prediction process, the prediction results can be corrected and adjusted according to the actual situation. For example, the prediction values ​​can be weighted averaged or otherwise reasonably corrected based on the historical operating rules of the energy storage power station, the current load conditions of the power grid, and other factors to further improve the reliability of the prediction.

[0138] Assume that the corrected predicted power value is P predict , the modified formula can be P′ predict =w 1 P predict +w 2 P hist +w 3 P grid , where w 1 ,w 2 ,w 3 is the weight coefficient determined based on experience or data analysis, P hist is the average power value under similar working conditions calculated based on historical data, P gridThe real-time power demand or limit value of the current power grid (if available); in this way, the prediction results are optimized by comprehensively considering multiple factors, so that the prediction values ​​are more in line with the actual operation of the energy storage power station, providing a more accurate basis for subsequent anti-backflow control decisions.

[0139] S4: Evaluate and adjust the prediction model.

[0140] The specific steps of evaluating and adjusting the prediction model are as follows:

[0141] S41: Evaluation index calculation:

[0142] During model training and prediction, the model is regularly evaluated using test set data.

[0143] In addition to the RMSE, MAE and R 2 In addition to the indicator, other evaluation indicators can also be calculated, such as the mean square error (MSE), which is calculated as follows: It is closely related to RMSE However, it can provide a more intuitive measure of the error size in some analyses.

[0144] In addition, the mean absolute percentage error (MAPE) is calculated as follows: It reflects the relative size of the prediction error in the form of a percentage, which makes it easier to compare the prediction accuracy of data of different sizes or different models.

[0145] S42: Model adjustment strategy:

[0146] According to the results of the evaluation indicators in step S41, if it is found that the model performance has decreased or the prediction error has increased, the model needs to be adjusted.

[0147] The adjustment strategy is to re-optimize the hyperparameters, increase the number of iterations of the TSSA algorithm, or adjust the parameters in the algorithm (such as λ in the Circle chaos map). 1 ,λ 2 ,λ 3 ) to find a better combination of hyperparameters. Another strategy is to increase the amount of training data. If new historical data is available, it can be added to the training set and the model can be retrained so that the model can learn more operating modes and features. The model structure can also be adjusted, such as increasing the number of hidden layers in the BLSTM layer or adjusting the depth of the decision tree in the RF model, but care should be taken to avoid overfitting. After adjusting the model, use the test set again to evaluate and observe whether the model performance has improved. Repeat this process until the model reaches a satisfactory performance level.

[0148] The present invention has been described by the above-mentioned relevant embodiments, however, the above-mentioned embodiments are only examples for implementing the present invention. It must be pointed out that the disclosed embodiments do not limit the scope of the present invention. On the contrary, changes and modifications made without departing from the spirit and scope of the present invention are all within the scope of patent protection of the present invention.

Claims

1. A method for preventing backflow in an energy storage power station based on a combined model, characterized in that: The specific steps of the method are as follows: S1: Data collection and preprocessing; The data include grid power data, environment and operation data; The preprocessing includes outlier processing and data filtering and smoothing processing; S2: Construct a bidirectional long short-term memory network-random forest prediction model; Including building bidirectional long short-term memory networks, random forest prediction, and TSSA optimization algorithms; S3: training and predicting the prediction model; Use the TSSA algorithm to optimize the hyperparameters of the bidirectional LSTM-random forest model to minimize the prediction error; S4: Evaluate and adjust the prediction model.

2. The energy storage power station backflow prevention control method based on the combined model according to claim 1, characterized in that: In step S1, the specific steps of data collection and preprocessing are as follows: S11: Grid power data collection: The grid power detection device collects grid power data in real time, including three-phase grid power and its direction information; at the same time, the energy storage converter obtains its actual output power and the operating parameters of the energy storage system, including battery voltage and current; S12: Environmental and operational data collection: Collect environmental data such as ambient temperature and humidity, as well as the operating status information of the energy storage power station, including charging and discharging mode and remaining battery capacity; S13: Outlier processing: Using a method based on statistical principles to detect and process outliers on the data collected in step S11 and step S12, and calculating the mean and standard deviation; S14: Data filtering and smoothing: The power data is filtered using a sliding average filtering algorithm.

3. The energy storage power station backflow prevention control method based on the combined model according to claim 1, characterized in that: In step S2, the specific steps of constructing a bidirectional long short-term memory network-random forest prediction model are as follows: S21: Building a bidirectional long short-term memory network: The bidirectional long short-term memory network is constructed by two LSTM layers, the forward LSTM layer and the backward LSTM layer. The bidirectional long short-term memory network extracts features that are important for power prediction based on factors such as past power change trends and possible future environmental impacts. S22: Construct random forest prediction: Since the operation data of energy storage power stations is complex, affected by multiple factors and has nonlinear relationships, multiple decision trees of random forests can learn and make decisions on different feature combinations. And since random forests are composed of multiple decision trees, the power of energy storage power stations is predicted based on the features extracted by the bidirectional long short-term memory network. S23: Construct TSSA optimization algorithm.

4. The energy storage power station backflow prevention control method based on the combined model according to claim 1, characterized in that: In step S3, the specific steps of training and predicting the prediction model are as follows: S31: Hyperparameter Optimization: The TSSA algorithm is used to optimize the hyperparameters in the bidirectional long short-term memory network-random forest model, including the number of hidden layers in the bidirectional long short-term memory network layer, the learning rate, and the number of decision trees in the random forest; the optimization goal is to minimize the prediction error; S32: Model training process: Input the preprocessed training set data into the bidirectional long short-term memory network, perform forward and backward propagation according to the calculation formula of the bidirectional long short-term memory network, and calculate the hidden layer state and output features; The output features of the bidirectional long short-term memory network are then used as the input of the random forest model. According to the training mechanism of the random forest, samples and features are randomly extracted to build a decision tree. In each iteration, the prediction error of the model on the training set is recalculated according to the TSSA algorithm update, that is, the hyperparameter combination, and the fitness value of the hyperparameter group is adjusted according to the error. S33: Real-time data input and processing: The pre-processed real-time data, including the current power, environmental data, and operating status data, is input into the trained TSSA bidirectional long short-term memory network and random forest model; the real-time data also needs to be normalized and pre-processed in the same way as the training data. S34: Feature extraction and prediction: The bidirectional long short-term memory network layer in the model first extracts features from the input real-time data, and calculates the hidden layer state and output feature vector according to the bidirectional long short-term memory network calculation formula mentioned above; then the extracted feature vector is input into the random forest model, and the random forest model makes predictions based on the trained decision tree.

5. The energy storage power station backflow prevention control method based on the combined model according to claim 1, characterized in that: In step S4, the specific steps of evaluating and adjusting the prediction model are as follows: S41: Evaluation index calculation: During model training and prediction, the model is regularly evaluated using test set data; S42: Model adjustment strategy: According to the results of the evaluation indicators in step S41, if it is found that the model performance has decreased or the prediction error has increased, the model needs to be adjusted.

6. The energy storage power station backflow prevention control method based on the combined model according to claim 1, characterized in that: In step S23, the specific steps of constructing the TSSA optimization algorithm are as follows: S231: Population initialization: Circle chaos mapping is used to initialize the sparrow population, improve the edge aggregation of sparrows, and improve the quality of the initial solution to better explore the feature space of energy storage power station related data; S232: Location Update: The location update includes the discoverer location update, the joiner location update, and the alerter location update; The discoverer position update is usually responsible for detecting, identifying or capturing certain key events. When the target location or related information changes, the discoverer will update the target location known to it; The joiner position update refers to following or joining other sparrows after discovering the target, usually entering the task execution area under the guidance of the discoverer; the joiner updates its own geographical location or status information according to the current task requirements or changes in the target location; The position update of the early warning device refers to the sparrows with early warning performance, which account for 10% to 20% of the sparrow population. The initial position is randomly generated by the system, and when danger is detected, its position is updated; S233: Add disturbance: The added disturbance refers to calculating the fitness of the population after the position is updated, finding the sparrow with the highest fitness value, and adopting the firefly disturbance strategy to update the optimal solution.

Images