An energy storage system for V2G application scenarios and its optimized control strategy

Abstract

The present invention discloses an energy storage system and its optimized control strategy for V2G application scenarios. The energy storage system is based on time-synchronized second pulses, integrating a combined prediction model of two time series prediction algorithms, ARIMA and LSTM, and is equipped with a double-layer optimization architecture of an MPC controller and a genetic algorithm, thus forming a multi-layer composite optimization control strategy. When the vehicle owner has random charging behavior, it is used to enable the three parties of EV, the energy storage system, and the power grid to use energy buffering to suppress supply-demand fluctuations and spatio-temporal mismatch according to the real-time data of the power grid. Through precise time synchronization, advanced prediction models, and optimization algorithms, the present invention can achieve a balance of interests among vehicle owners, energy storage operators, and the power grid while considering the random charging behavior of vehicle owners.

Description

Technical Field

[0001] The present invention relates to an energy storage system in a power system, and particularly to an energy storage system and an optimized control strategy in an application scenario of an electric vehicle providing energy storage to the power grid (V2G). Technical Background

[0002] V2G (Vehicle-to-Grid) technology refers to the interaction between electric vehicles (EVs) and the power grid. Through the bidirectional energy flow between the batteries of electric vehicles and the power grid, the dynamic balance of energy and the optimization of resources are achieved. Vehicles can not only charge from the power grid but also feed back electrical energy to the power grid at high electricity prices, realizing the bidirectional flow of energy, and is gradually becoming an effective energy management method. However, with the large-scale application of electric vehicles (EVs) and renewable energy, the problems of load management and power quality faced by the power grid are becoming increasingly serious. At the same time, with the increase in the number of EVs, the randomness of vehicle owners' charging behaviors brings many challenges and complexity to the bidirectional and orderly interaction of the power grid, and affects the interests of vehicle owners, energy storage operators, the power grid, and the stability of the power grid.

[0003] As a buffer between EVs and renewable energy and the power grid, the energy storage system needs to smooth the power output curve of the power source side by energy storage on the one hand, and on the other hand, adjust it by a flexible and adjustable load at the power consumption end. At the same time, considering factors such as time-varying electricity prices and charging demand forecasting (charging time series characteristics), how to ensure the operation efficiency and economic benefits of the overall power system, so that the energy storage system can provide such an ability, that is, to suppress the supply-demand fluctuations and spatio-temporal mismatches with energy buffering according to the real-time data of the power grid, which is one of the technical contents that the applicant has been committed to researching and developing.

[0004] Therefore, in the V2G application scenario, as a buffer between EVs and renewable energy and the power grid, how to ensure the balance of interests among vehicle owners, energy storage operators, and the power grid, and consider the randomness of vehicle owners' charging behaviors, is a realistic problem faced by the energy storage system currently. Summary of the Invention

[0005] The present invention provides an energy storage system and its optimized control strategy for V2G application scenarios. Based on the time synchronization second pulse technology, a prediction model integrating ARIMA (Autoregressive Integrated Moving Average) and LSTM (Long Short-Term Memory), combined with a multi-layer composite optimization control strategy of MPC (Model Predictive Control) and GA (Genetic Algorithm), this control strategy can achieve the interest balance among vehicle owners, energy storage operators, and the power grid by means of precise time synchronization, advanced prediction models, and optimization algorithms, considering the random charging behavior of vehicle owners.

[0006] The technical solution to achieve the above object is: An energy storage system for V2G application scenarios, characterized in that: the energy storage system is based on the time synchronization second pulse, integrating a combined prediction model of two time series prediction algorithms of ARIMA and LSTM, and equipped with a double-layer optimization architecture of an MPC controller and a genetic algorithm, thus forming a multi-layer composite optimization control strategy, so that when vehicle owners have random charging behaviors, the EV, the energy storage system, and the power grid can use energy buffering to suppress the supply-demand fluctuations and spatio-temporal mismatches according to the real-time data of the power grid.

[0007] In the above-mentioned energy storage system for V2G application scenarios, the time synchronization second pulse uses a high-precision timing reference to ensure that the data acquisition and control operations of each node of the energy storage system are carried out under the same time reference, ensuring the timing consistency of the entire system's data acquisition, processing, and control decisions.

[0008] The system hardware of the time synchronization second pulse includes: a BD / GPS dual-mode receiver, a high-precision clock synchronization module, and a disciplined clock sub-module.

[0009] The BD / GPS dual-mode receiver receives the standard time signal transmitted by the satellite, and generates a 1PPS signal through the processing of the high-precision clock synchronization module; the BD / GPS dual-mode receiver is connected to the high-precision clock synchronization module and deployed in the energy storage system. The time synchronization interface of the high-precision clock synchronization module outputs the 1PPS signal to the corresponding interface of the energy control main board of the energy storage system to ensure that all control modules of the energy storage system maintain clock synchronization under a unified and high-precision time reference.

[0010] The disciplined clock sub-module is deployed in the V2G bidirectional charging pile and connected to its control module, and is also connected to the time synchronization interface of the high-precision clock synchronization module of the energy storage system. It is calibrated and disciplined by receiving the 1PPS signal provided by it to achieve clock synchronization consistent with the energy storage system.

[0011] In the above energy storage system for V2G application scenarios, the ARIMA prediction model is used to analyze and predict linear time series data in time series, and identify and capture the trends and seasonal variations of the data through the historical charging data of electric vehicles;

[0012] The LSTM prediction model is used to learn and predict time series data with non-linear and long-term dependency relationships in time series, and can effectively capture the randomness of vehicle owners' charging behaviors, including changes with factors such as time, temperature, and working day / holiday;

[0013] The combined prediction model of the two prediction algorithms can provide more accurate charging load predictions for the energy storage system, making the energy storage system more intelligent in the face of random and volatile demand changes at different time scales. It can synchronize the system operation with the grid load fluctuations, accurately optimize the charging and discharging strategies, use the ARIMA model to handle short-term linear dependencies, and use the LSTM model to capture long-term dependencies and non-linear patterns such as complex patterns. The role of the time synchronization second pulse in this combined prediction model is to ensure that the time scales of all time series data are synchronized with the positions at the actual event occurrence, thereby improving the accuracy and reliability of the prediction.

[0014] In the above energy storage system for V2G application scenarios, in the two-layer optimization architecture of the MPC controller and genetic algorithm, where:

[0015] The MPC controller generates an optimal charging / discharging strategy in real time according to the results provided by the combined prediction model of ARIMA and LSTM to minimize the operating cost and meet the grid demand;

[0016] The genetic algorithm is used to optimize the MPC controller parameters to ensure finding the best control strategy in the V2G environment, that is: taking the control strategy generated by the MPC controller as the initial population of the genetic algorithm, randomly generating an initial parameter set, calculating the appropriate fitness function for each individual, evaluating the advantages and disadvantages of the economic benefits of vehicle owners, energy storage operators, and the grid, and performing selection, crossover, and mutation genetic operations on individuals with high fitness to generate a new generation of population, that is, the optimal control strategy that satisfies all three of EV, energy storage system, and the grid.

[0017] In the above energy storage system for V2G application scenarios, the energy storage system receives a time synchronization signal at the beginning of each whole second, compares and calibrates this signal with the clocks of each control module inside the system to ensure that all control modules inside the system operate on the same time standard. Within each synchronization cycle, the high-precision clock synchronization module 1PPS updates the clocks of each disciplined clock sub-module for time calibration and deviation correction. The specific algorithm is:

[0018] ΔTi = T BD / GPS -T local,i

[0019]

[0020] Where: T BD / GPS represents the standard time of BD / GPS received by the high-precision clock synchronization module, that is, 1PPS, and T local,i is the local clock time of the i-th subsystem, and ΔT i is the measured time deviation, is the corrected local time.

[0021] The present invention also provides an optimized control strategy for an energy storage system, including the following steps:

[0022] Step 1, receive the BD / GPS time synchronization signal and perform second pulse time synchronization;

[0023] Step 2, collect charge and discharge time series data;

[0024] Step 3, combined prediction model training, use the ARIMA model to process the short-term linear relationship of charging behavior data, and then use these prediction results as inputs combined with other features, including temperature and date type, and use the LSTM model to capture and predict long-term, non-linear patterns;

[0025] Step 4, real-time load prediction, gradually reduce the prediction error by iteratively updating the weights and biases to achieve more accurate prediction of future demand;

[0026] Step 5, MPC control strategy generation, the MPC controller generates an optimal charge / discharge control strategy for the energy storage system based on the prediction results of the time series relationship generated by the ARIMA and LSTM models, and can be adjusted according to real-time data to ensure the stable operation and optimized control of the system in a dynamic environment;

[0027] Step 6, genetic algorithm optimization, the genetic algorithm optimizes the parameters of the MPC controller to avoid falling into local optima, and through the iterative evolution of the genetic algorithm, realizes the adaptive adjustment of the control strategy and improves the control effect;

[0028] Step 7, generate charge / discharge strategies and execute charge / discharge strategies.

[0029] In the above optimized control strategy for an energy storage system, the ARIMA model in step 3 is specifically as follows:

[0030] Short-term linear prediction using the ARIMA model: The ARIMA model consists of three parameters p, d, and q. p is the autoregressive order; d is the differencing order; q is the moving average order. The parameters d, p, and q of the ARIMA model are determined through stationarity tests, calculations, and observing the autocorrelation function and partial autocorrelation function graphs.

[0031] d is determined by the stationarity test, the ADF test, that is, using the ADF test to determine the stationarity of the time series.

[0032] ADF test formula: Δy t = α + βt + γy t-1 + δ1Δy t-1 + δ2Δy t-2 +…+ δ k Δy t-k + ε t

[0033] Where, Δy t is the differenced term, α is the constant, βt is the time trend term, γy t-1 is the lag term, δ i Δy t-i is the lagged differenced term, ε t is the random error term.

[0034] The autoregressive order p and the moving average order q are determined by observing the autocorrelation function ACF and the partial autocorrelation function PACF.

[0035] ACF formula:

[0036] Where: ρ k is the autocorrelation coefficient at lag k; y t is the value of the time series at time t; y t-k is the value of the time series at time t - k, representing the value k time points forward from time t. is the mean of the time series; T is the total length of the time series.

[0037] PACF formula:

[0038] Where: φ kk is the partial autocorrelation coefficient at lag k, representing the autocorrelation of the sequence excluding the intermediate lag terms at lag k; ρ k is the autocorrelation coefficient at lag k, representing the linear correlation between the current moment and the value at lag k moments of the sequence; φ k-1,i is the partial autocorrelation number at lag k - 1, which is the intermediate term to be considered when calculating the partial autocorrelation coefficient at the current lag k; ρ k-iis the autocorrelation coefficient at lag k - i, which is used to adjust the value of the partial autocorrelation coefficient at the current lag k.

[0039] ARIMA model construction and prediction: Input the estimated parameter values into the ARIMA model to establish an ARIMA(p, d, q) model. Through the combined action of differencing, regression, and moving average, short-term linear prediction can be performed on time series data, thus being used for predicting charging demand.

[0040] In the above optimization and control strategy for an energy storage system, the LSTM model in step 3 is specifically as follows:

[0041] Using LSTM for long-term non-linear prediction: The LSTM model controls the flow of information through three "gate" structures, namely: the forget gate, the input gate, and the output gate, thus effectively preserving long-term dependencies.

[0042] First, the forget gate f t , which is used to determine what information to discard from the cell state.

[0043] f t = σ(W f · [h t-1 , X t + b f )

[0044] Where: W f is the weight matrix of the forget gate; [h t-1 , X t is the concatenation of the hidden state at the previous time point and the current input; b f is the bias of the forget gate; σ is the sigmoid activation function, which outputs a scalar between 0 and 1, representing the retention ratio.

[0045] Second, the input gate i t and the candidate cell state

[0046] i t = σ(W i · [h t-1 , X t + b i ), function: Determine what new information to store in the cell state.

[0047]

[0048] Where W i , W c are the weight matrices of the input gate; b i , b cis the bias term of the input gate; tanh is the hyperbolic tangent activation function, which is used to generate the candidate cell state, and the output range is between -1 and 1;

[0049] Third, the cell state update Y t , which is used to update the cell state, combining the forgotten and newly input information,

[0050]

[0051] where Y t is the cell state at the current time point; Y t-1 is the cell state at the previous time point;

[0052] Fourth, the output gate o t and the hidden state h t :

[0053] o t = σ(W o ·[h t-1 , X t +b o ), which is used to determine what information to output as the output at this time point, h t = o t *tanh(Y t ), that is, finally, the actual demand prediction is output through the transfer function

[0054] where: W o is the weight matrix of the output gate; b o is the bias term of the output gate.

[0055] In the above optimization and control strategy of an energy storage system, the calculation steps of the LSTM model in step 3 are as follows:

[0056] Step A, initial step:

[0057] Collect the features related to the charging demand, including: temperature, time, and holidays, and create the feature set X t , determine the hidden layer, which contains a number of LSTM units, and each unit captures the time dependence in the data,

[0058] Step B, data processing:

[0059] Combine the prediction result C t+1 of the ARIMA model with other features X t , including the data from t - n to t, to construct the training set [X t-n , …, X t , C t+1 of the LSTM model;

[0060] Step C, LSTM model training:

[0061] The training of the LSTM model is an iterative process that uses historical time series data to train the LSTM network through steps such as forward propagation, loss function, backpropagation, and parameter update to determine the weight matrix and bias values;

[0062] Among them, the goal of the loss function L is to minimize the error between the predicted output and the actual demand, and the mean square error MSE is adopted:

[0063]

[0064] Among them: Z t+1 is the actual charging demand value, is a predicted charging demand value output by the LSTM based on the current weights and biases at each step of the iteration, and N is the number of samples.

[0065] Backpropagation is used to calculate the gradient of the loss function with respect to each model parameter (weights and biases). The gradient represents the rate of increase of the loss function near the current value of the parameter and is used to update the weights to reduce the error. The parameter update is calculated using the gradient descent method:

[0066]

[0067] Among them, α is the learning rate, and L is the loss function. and are the gradients of the weight matrix and bias respectively. W old and b old are the unified expressions of the weight matrix and bias values of the three "gates" respectively. and are the updated weight matrix and bias values respectively, which are updated and adjusted according to the gradient of the loss function in each training iteration, so that the model can better learn and capture the underlying patterns of the data;

[0068] Step D, prediction:

[0069] Input the training set of the LSTM model [X t-n , …, X t , C t+1 into the LSTM formula after updating the weight matrix and bias values and calculate the final prediction at time t + 1 through the transfer function.

[0070] In the above optimization and control strategy of an energy storage system, in step 5, the MPC controller controls through the following process:

[0071] First, the dynamic behavior of the energy storage system can be described by a linear state space model:

[0072]

[0073] The output equation is: y k = Dx k + v k

[0074] where x k is the system state vector at time k; u k is the control input vector at time k; is the predicted electricity demand at time k+1 predicted by the ARIMA and LSTM models; A is the state transition matrix; B is the control matrix; C is the input matrix; y k is the system output vector at time k; w k is the process noise; D is the output matrix; v k is the measurement noise;

[0075] Second, the goal of MPC is to optimize the control input u k within the prediction horizon of the ARIMA and LSTM models, k so that the system state x ref tracks the desired state x

[0076]

[0077] where is the sequence of predicted state vectors from time k+1 to k+N; i = 1, 2, 3…N, is the predicted electricity demand at time k+i, predicted by the ARIMA and LSTM models; the model calculates the future state recursively, where each predicted state x k+i is based on the current state x k+i-1 , the control input u k+i-1 and the predicted demand calculated,

[0078] Third, the goal of the MPC controller is to minimize the operating cost and meet the grid demand, and the objective function is expressed as:

[0079]

[0080] where the vector set of all control inputs within the prediction horizon is expressed as: x i - x ref represents the state deviation at time i, that is, the gap between the current state and the desired state; (x i - x ref )T Q(x i - x ref ) represents the weighted quadratic norm of the state deviation, that is, the gap between the current state and the desired state; Q represents the state deviation weighting matrix; u i represents the control input vector at time i; represents the weighted quadratic norm of the control input, that is, the magnitude of the control input; R represents the control input weighting matrix; the control input u i and the state x i need to satisfy certain constraint conditions:

[0081]

[0082] The beneficial effects of the present invention are as follows: By applying the time synchronization second pulse technology to the energy storage system, the technical solution of the present invention ensures a high degree of synchronization inside and outside the system, improves the overall operation efficiency, and integrates the ARIMA and LSTM prediction models with a unified time reference. It can not only capture the linear trend of the data but also process complex non-linear features, improving the accuracy of load prediction. Then, by combining MPC and genetic algorithm based on the prediction results, the on-time optimal control of the system is realized, enhancing the effectiveness and stability of the control strategy, ensuring that the control signal can be delivered and executed at the correct moment, thereby achieving higher prediction accuracy and adaptive control of the energy storage system. At the same time, the second pulse technology ensures that all devices operate within the same time frame in complex application scenarios with a large number of devices, enabling the coordination and synchronization of the overall power system. Compared with the prior art, this solution significantly improves the operation efficiency and stability of the energy storage system. By optimizing the control strategy, it reduces the overcharge and over-discharge behavior of the energy storage system battery, extends the battery life, and optimizes the energy management strategy, reducing energy loss and operating costs, improving economic benefits, and thus achieving better performance in a dynamically changing market environment and participant behavior, providing more resilient support for the power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 is the system architecture diagram of the energy storage system of the present invention;

[0084] Figure 2 is the flow chart of the optimized management and control strategy of the energy storage system of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0085] The present invention will be further described below with reference to the accompanying drawings.

[0086] 1. The technical solution of the energy storage system of the present invention is as follows (see Figure 1 ):

[0087] 1. Time synchronization second pulse technology

[0088] The time synchronization second pulse technology plays a series connection role in the solution of the present invention. By using a high-precision timing reference, it ensures that the data acquisition and control operations of each node in the system are carried out under the same time reference, ensuring the timing consistency of data acquisition, processing, and control decision-making of the entire system, and laying a foundation for implementing efficient prediction and optimal control.

[0089] 2. ARIMA and LSTM Combined Prediction Model

[0090] In the present invention, the energy storage system aims to balance the load demand of the power grid. This system needs to predict the fluctuations in the power grid to optimize storage and discharge operations. A combined prediction model that combines two time series prediction algorithms, ARIMA and LSTM, is proposed in the solution for charging behavior prediction:

[0091] The ARIMA model is used to analyze and predict linear time series data in a time series. It identifies and captures the trends and seasonal variations in the data well through the historical charging data of electric vehicles.

[0092] The LSTM model is used to learn and predict time series data with non-linear and long-term dependence relationships in a time series. It can effectively capture the randomness of the charging behavior of vehicle owners, such as changes with factors like time, temperature, weekday / holiday, etc.

[0093] The combination of the two models can provide a more accurate charging load prediction for the energy storage system, making the energy storage system more intelligent in the face of randomness and volatility demand changes at different time scales. It can synchronize the system operation with the power grid load fluctuations, accurately optimize the charging and discharging strategies, and bring double economic and environmental benefits to the power system. The ARIMA model is used to handle short-term linear dependencies (seasonality and trends), and the LSTM model is used to capture long-term dependencies and non-linear patterns such as complex patterns. For example, the ARIMA model is used to obtain preliminary charging demand prediction values for trends and seasonality, and then this value and other relevant features (such as temperature, date type, etc.) are input into the LSTM model to learn and further iteratively refine these data to obtain a more accurate final prediction. The role of the time synchronization second pulse technology in this model is to ensure that the time stamps of all time series data are synchronized with the positions at the time of actual events, thereby improving the accuracy and reliability of the prediction.

[0094] 3. MPC Controller and Genetic Algorithm Double-Layer Optimization Architecture

[0095] The MPC controller plays a core role in the solution of the present invention. It generates an optimal charging / discharging strategy in real time according to the results provided by the ARIMA and LSTM combined prediction model to minimize the operating cost while meeting the power grid demand (objective function), achieving the dual goals of economy and security.

[0096] The GA (Genetic Algorithm) is used to optimize the parameters of the MPC (Model Predictive Control) controller to ensure finding the optimal control strategy in a complex and uncertain V2G environment. By taking the control strategy generated by the MPC as the initial population of the genetic algorithm, randomly generating an initial parameter set, and calculating the appropriate fitness function for each individual (in this case, the fitness function is defined as the opposite of the objective function, that is, when the objective function is minimized, the fitness is maximized), the economic benefits of the vehicle owner, energy storage operator, and power grid are evaluated for their advantages and disadvantages. Genetic operations such as selection, crossover, and mutation are performed on individuals with high fitness to generate a new generation of population, that is, the optimal control strategy that meets the needs of multiple stakeholders.

[0097] The combination of the two algorithms can provide the optimal control strategy for the energy storage system in the V2G scenario, enabling the energy storage system to achieve the best balance between economic benefits and power supply-demand balance. The MPC uses a dynamic model to predict future outputs and calculates the optimal control decision. The GA is used to search for the optimal MPC configuration to achieve the efficient operation of the system in the face of uncertainty and dynamic market conditions. The precise time stamp provided by the time synchronization second pulse technology ensures that the control signals optimized and output by the MPC controller and the genetic algorithm are delivered and executed at the correct moment, enabling the coordination and synchronization of the overall system.

[0098] 4. Implementation Steps

[0099] - Use the time synchronization second pulse technology to synchronize the data flow in the V2G system.

[0100] - Use historical data to train the combined ARIMA and LSTM models.

[0101] - Apply the trained model for real-time load forecasting.

[0102] - Use the MPC and genetic algorithm to generate an optimized control strategy.

[0103] - Apply the control strategy to manage the charging / discharging process of the energy storage system.

[0104] The above steps can be referred to Figure 2 the flowchart.

[0105] II. Specific Implementation of the Energy Storage System of the Present Invention

[0106] 1. Implementation of the Time Synchronization Second Pulse Technology

[0107] The time synchronization second pulse technology is the key to ensuring accurate data synchronization in the V2G energy storage system. This technology uses a high-precision clock synchronization signal to coordinate the time of all devices within the system. This is crucial for accurately implementing data collection, processing, and control decisions.

[0108] In the time synchronization second pulse system hardware part of the present invention's solution, it includes: a BD / GPS dual-mode receiver, a high-precision clock synchronization module, and a disciplined clock sub-module. The BD / GPS dual-mode receiver receives the standard time signal transmitted by the satellite, and generates a 1PPS signal through the processing of the high-precision clock synchronization module. The BD / GPS dual-mode receiver is connected to the high-precision clock synchronization module and deployed in an energy storage system with an energy storage cabinet as the main body. The time synchronization interface of the high-precision clock synchronization module outputs the 1PPS signal to the corresponding interface of the energy control main board of the energy storage system to ensure that all control modules in the energy storage system maintain clock synchronization under a unified and high-precision time reference. The disciplined clock sub-module is deployed in the V2G bi-directional charging pile, connected to its control module, and connected to the time synchronization interface of the high-precision clock synchronization module of the energy storage system. It is calibrated and disciplined by receiving the 1PPS signal provided by it to achieve clock synchronization consistent with the energy storage system.

[0109] The entire energy storage system receives a time synchronization signal at the beginning of each whole second, compares and calibrates this signal with the clocks of each control module inside the system to ensure that all control modules in the system operate on the same time standard. During each synchronization cycle, the high-precision clock synchronization module updates the clock of each disciplined clock subsystem with 1PPS, and performs time calibration and deviation correction algorithms as follows:

[0110] ΔT i =T BD / GPS -T local,i

[0111]

[0112] Where, T BD / GPS represents the standard time of BD / GPS received by the high-precision clock synchronization module, that is, 1PPS. T local,i is the local clock time of the i-th subsystem, and ΔT i is the measured time deviation, is the corrected local time.

[0113] 2. Implementation of the ARIMA and LSTM combined prediction model

[0114] When the energy storage system predicts future electricity consumption behaviors, it needs to use a prediction function with time as the independent variable, that is, the time parameters of each control part of the energy storage system all use as the benchmark. In the present invention's solution, the ARIMA and LSTM combined prediction model utilizes their respective advantages to improve the accuracy of charging load prediction. First, the ARIMA model is used to process the short-term linear relationship of charging behavior data, and then these prediction results are used as inputs and combined with other features, such as temperature, date type, etc. The LSTM model is used to capture and predict long-term and possibly more complex non-linear patterns.

[0115] Step 1: Data Preparation and Preprocessing

[0116] 1. Data collection: Obtain historical charging data, including the charging time and charging volume of electric vehicles, as well as data on other external factors that may affect charging demand, such as temperature, day of the week, holidays, time-of-use electricity price, etc.

[0117] 2. Data cleaning: Handle missing values, remove outliers, and standardize the data to eliminate the impact of dimensions.

[0118] 3. Feature engineering: Construct new features, extract weekdays and weekends from dates, extract peak and off-peak periods from time-of-use electricity prices, etc.

[0119] Step 2: Use ARIMA for Short-Term Linear Prediction

[0120] After completing Step 1, use ARIMA for short-term linear prediction. The ARIMA model captures the characteristics of the time series by describing the autoregressive (AR), differencing (I), and moving average (MA) parts of the time series. The ARIMA(p,d,q) model consists of three main parameters: p (the order of autoregression, representing how many past values it depends on), d (the order of differencing, the number of differencing operations required to make the time series stationary), and q (the order of moving average, representing how many past error terms it depends on). The calculation process is as follows:

[0121] 1. Model selection and parameter estimation

[0122] Build the ARIMA model in the form of ARIMA(p,d,q), where

[0123] d is determined by the stationarity test, the ADF test (Augmented Dickey-Fuller Test), that is, use the ADF test to determine the stationarity of the time series. If the time series is not stationary, perform differencing operations to make it stationary. If the statistic of the ADF test is less than the significance level (usually 0.05), then the time series can be considered stationary, otherwise perform differencing operations.

[0124] ADF test formula: Δy t =α + βt + γy t-1 + δ1Δy t-1 + δ2Δy t-2 +…+ δ k Δy t-k + ε t

[0125] where, Δy t is the differenced term, α is the constant, βt is the time trend term, γy t-1 is the lag term, δi Δy t-i is the lag difference term, and ε t is the random error term.

[0126] Differencing operation: If the original sequence is non-stationary (i.e., the null hypothesis cannot be rejected), then perform the differencing operation and then re-conduct the ADF test. Perform the differencing operation successively until the sequence becomes stationary. The number of differencing operations is the parameter d.

[0127] p and q are determined by the autocorrelation (ACF) and partial autocorrelation (PACF):

[0128] The autoregressive order p and the moving average order q can be determined by observing the autocorrelation function (ACF) and the partial autocorrelation function (PACF).

[0129] ACF is used to describe the autocorrelation of time series data at different lags k, that is, the correlation between a sequence y t and its own lagged value y t-k . ACF helps to identify the seasonality and periodicity of the sequence by plotting the autocorrelation coefficients at different lag values.

[0130] ACF formula:

[0131] where: ρ k is the autocorrelation coefficient at lag k; y t is the value of the time series at time t; y t-k is the value of the time series at time t - k, representing the value k time points forward from time t. is the mean of the time series; T is the total length of the time series.

[0132] PACF is used to describe the partial autocorrelation of the measured time series at different lags k, aiming to exclude the influence of intermediate lag values to determine the direct correlation. PACF is used to identify the lag order of direct correlation in the time series, that is, the order of the AR model.

[0133] PACF formula:

[0134] where: φ kk is the partial autocorrelation coefficient at lag k, representing the autocorrelation of the sequence excluding intermediate lag terms at lag k; ρ k is the autocorrelation coefficient at lag k, representing the linear correlation between the sequence at the current moment and the value at lag k time points; φ k-1,i is the partial autocorrelation number at lag k - 1, which is the intermediate term to be considered when calculating the partial autocorrelation coefficient at the current lag k; ρ k-i is the autocorrelation coefficient at lag k - i, which is used to adjust the value of the partial autocorrelation coefficient at the current lag k.

[0135] Calculate the ACF and PACF for each lag k according to the above formula

[0136] Determine the autoregressive order p:

[0137] Characteristics of the autoregressive model (AR): The ACF plot decays gradually; the PACF plot truncates after the order p (i.e., the partial autocorrelation coefficients after the significant lag order are significantly zero. If it becomes zero after being significantly non-zero at lag k, then p = k).

[0138] By observing the PACF plot, when the PACF truncates at a certain lag order p, this lag order is the order p of the AR model.

[0139] Determine the moving average order q:

[0140] Characteristics of the moving average model (MA): The PACF plot decays gradually; the ACF plot truncates after the order q (i.e., the autocorrelation coefficients after the significant lag order are significantly zero. If it becomes zero after being significantly non-zero at lag k, then q = k).

[0141] By observing the ACF plot, when the ACF truncates at a certain lag order q, this lag order is the order q of the MA model.

[0142] In summary, the parameters d, p, and q of the ARIMA model are determined through the stationarity test (ADF test) and calculating and observing the autocorrelation function (ACF) and partial autocorrelation function (PACF) plots. Through these steps, the model parameters suitable for time series data can be reasonably determined and applied to prediction.

[0143] 2. ARIMA Model Construction and Prediction

[0144] Input the estimated parameter values into the ARIMA model to establish the ARIMA(p, d, q) model for short-term linear prediction.

[0145] Predict the charging demand C at time point t + 1 based on the current and historical data t+1 , establish the model formula:

[0146]

[0147] where C t is the value of the time series at time point t, L is the lag operator, which shifts the time series data backward by k time points, i.e., L k C t = C t-k , φ i is the autoregressive coefficient (representing the linear relationship between the current value and its past values in the time series), i represents the lag number, θi is the moving average coefficient (indicating the magnitude of the influence of past prediction errors on the current prediction error), ε t is the white noise error term, representing the random error at time point t.

[0148] The mathematical expression of the AR part is:

[0149]

[0150] The mathematical expression of the I part is:

[0151] I: (1 - L) d C t

[0152] The mathematical expression of the MA part is:

[0153]

[0154] According to the above formulas of each part, calculations and interpretations can be performed according to the ARIMA(1, 1, 1) model. First, a differencing operation needs to be carried out, that is, perform d = 1 time of differencing on the time series data {C t} to make it stationary, and obtain the new time series data {ΔC t} after differencing. ΔC t =(1 - L)C t =C t -C t-1 . Then, based on the new time series data after differencing, an ARIMA model fitting is carried out. The solution of the present invention uses the Yule-Walker equation to solve the autoregressive coefficient φ1, and the maximum likelihood estimation method is used to solve the moving average coefficient θ1. Finally, according to the obtained φ1 and θ1, the ARIMA(p, d, q) model formula is iteratively updated to construct a prediction model for practical application to predict future time points. According to the ARIMA(1, 1, 1) model, the predicted difference value ΔC t+1 =φ1ΔC t +ε t+1 +θ1ε t can be obtained at the next moment t + 1, where it is assumed that ε t and ε t+1 are values close to zero. Then, the predicted difference value is converted back to the predicted value of the original sequence, that is, the value C t+1 =C t +ΔC t+1 at the next moment is predicted.

[0155] In summary, through the combined action of differencing, regression, and moving average, the ARIMA model can model and predict time series data, and thus be used for predicting charging demand.

[0156] Step 3: Use LSTM for long-term non-linear prediction

[0157] In the solution of the present invention, the charging demand prediction is divided into two steps, that is, the charging demand prediction value C output according to the above ARIMA model t+ 1 is combined with other relevant features (such as temperature, date type, and user behavior trend, etc.) and input into the LSTM model for training to obtain the final charging demand prediction value

[0158] The LSTM model controls the flow of information by introducing three "gate" structures (forget gate, input gate, and output gate), so as to effectively preserve long-term dependencies. The model formula is as follows:

[0159] 1. Forget gate f t , Function: Determine what information to discard from the cell state.

[0160] f t =σ(W f ·[h t-1 ,X t +b f ).

[0161] Where W f is the weight matrix of the forget gate; [h t-1 ,X t is the connection of the hidden state at the previous time point and the current input; b f is the bias of the forget gate; σ is the sigmoid activation function, which outputs a scalar from 0 to 1, indicating the retention ratio.

[0162] 2. Input gate i t and candidate cell state

[0163] i t =σ(W i ·[h t-1 ,X t +b i ), Function: Determine what new information to store in the cell state;

[0164]

[0165] Where W i ,W c are the weight matrices of the input gate; b i ,b c are the bias terms of the input gate; tanh is the hyperbolic tangent activation function, which is used to generate the candidate cell state, and the output range is between -1 and 1.

[0166] 3. Cell state update Y t, Function: Update the cell state, combining forgotten and newly input information.

[0167]

[0168] Where Y t is the cell state at the current time point; Y t-1 is the cell state at the previous time point.

[0169] 4. Output gate o t and hidden state h t :

[0170] o t = σ(W o · [h t-1 , X t + b o ), Function: Determine what information to output as the output at this time point. h t = o t * tanh(Y t ), that is, finally output the actual demand prediction through a transfer function (such as a linear function, a and b are linear transformation parameters, obtained by training using linear regression to minimize the mean squared error between the actual output and the predicted output)

[0171] Where W o is the weight matrix of the output gate; b o is the bias term of the output gate.

[0172] The specific calculation steps of the LSTM model are as follows:

[0173] 1. Initial step:

[0174] Collect features related to the charging demand: temperature, time (hours, day of the week), holidays, etc., and create a feature set X t . Determine the hidden layer, which contains a number of LSTM cells, and each cell captures the temporal dependence in the data.

[0175] 2. Data processing:

[0176] Combine the prediction result C t+1 of the ARIMA model with other features X t (including data from t - n to t), and construct the training set [X t-n , …, X t , C t+1 of the LSTM model.

[0177] 3. LSTM model training:

[0178] The training of the LSTM model is an iterative process that uses historical time series data to train the LSTM network through steps such as forward propagation, loss function, backpropagation, and parameter update to determine the weight matrix and bias values.

[0179] Among them, the goal of the loss function L is to minimize the error between the predicted output and the actual demand, and the mean squared error (MSE) is adopted:

[0180] where Z t+1 is the actual charging demand value, is a predicted charging demand value output by the LSTM based on the current weights and biases at each step of the iteration using the training set, and N is the number of samples.

[0181] Backpropagation is used to calculate the gradient of the loss function with respect to each model parameter (weights and biases). The gradient represents the rate of increase of the loss function near the current values of the parameters and is used to update the weights to reduce the error. Parameter update is calculated using the gradient descent method:

[0182]

[0183] where α is the learning rate, L is the loss function, and are the gradients of the weight matrix and bias respectively, and W old and b old are the unified expressions of the weight matrices and bias values of the three "gates" respectively. and are the updated weight matrix and bias values respectively, which are updated and adjusted according to the gradient of the loss function in each training iteration, so that the model can better learn and capture the underlying patterns of the data.

[0184] 4. Prediction:

[0185] Input the training set [X t-n , …, X t , C t+1 of the LSTM model into the LSTM formula after updating the weight matrix and bias values and calculate the final prediction at time t + 1 through the transfer function

[0186] In summary, the solution of the present invention combines the ARIMA model and the LSTM model, and uses the ability of the ARIMA model to capture linear and short-term dependency features and the LSTM model to capture complex long-term dependency features, thereby solving the problem that a single model is difficult to capture complex patterns. By iteratively updating weights and biases to gradually reduce prediction errors, a more accurate prediction of future demand can be achieved. In practical applications, the combined model can more accurately predict the charging demand load of electric vehicles in the future, and can optimize the scheduling and operation strategies of the energy storage system by scheduling energy storage resources in advance, thereby improving its operating efficiency and reliability, and reducing operating costs.

[0187] 3. Implementation of the MPC controller and genetic algorithm two-layer optimization architecture

[0188] In the solution of the present invention, the MPC controller generates the optimal charge / discharge control strategy for the energy storage system based on the prediction results of the time series relationship generated by the ARIMA and LSTM models, and can be adjusted according to real-time data to ensure the stable operation and optimized control of the system in a dynamic environment. The genetic algorithm optimizes the parameters of the MPC controller to avoid falling into the local optimum. Through the iterative evolution of the genetic algorithm, the control strategy is adaptively adjusted to improve the control effect. In the entire energy storage system, prediction and control are based on time series data, and time synchronization ensures the accuracy and real-time synchronization of data collection, analysis, control signal transmission and execution.

[0189] The implementation process is as follows:

[0190] The MPC controller adjusts the control strategy in real time through online rolling optimization to minimize operating costs and meet grid demand, while ensuring that the charge level soc of the energy storage system is within a safe operating range. In the scheme of the present invention, the dynamic behavior of the energy storage system can be described by a linear state space model:

[0191] The output equation is: k =Dx k +v k

[0192] Among them, x k is the system state vector at time k, representing the current state of the energy storage system, such as the charge level and charge / discharge rate of the energy storage battery; u k is the control input vector at time k, representing the current control action on the energy storage system, such as charging and discharging power; (Equivalent to above ) is the predicted electricity demand at time k+1 obtained from the ARIMA and LSTM models; A is the state transition matrix, describing the autoregressive relationship of the system state (such as the relationship between battery charge and charge / discharge rate); B is the control matrix, describing the impact of the control input on the system state (such as the impact of the control input on charge and rate); C is the input matrix, describing the impact of the predicted charge / discharge demand on the system state (A, B, and C can be determined by regression analysis of experimental data or historical data); y k is the system output vector at time k, such as including the monitored battery state, etc., which is used for feedback control and state observation during the control process of MPC to help adjust the control input of the system; w k is the process noise, representing the uncertainty of the system, which can be assumed to be zero-mean Gaussian white noise; D is the output matrix, describing the impact of the system state on the output; v k is the measurement noise, representing the uncertainty during the measurement process, which can be assumed to be zero-mean Gaussian white noise.

[0193] Continuing, the goal of MPC is to optimize the control input u k within the prediction time domain of the ARIMA and LSTM models, such that the system state x k tracks the desired state x ref (representing the desired system state, such as the target charge and charge / discharge rate). The length of the prediction time domain is N, then the control model of the predicted state and control input within the next N time instants can be expressed as:

[0194]

[0195] where is the sequence of predicted state vectors from time k+1 to k+N;

[0196] (i = 1, 2, 3…N) is the predicted electricity demand at time k+i, obtained from the ARIMA and LSTM models. The model calculates the future state through recursion, where each predicted state x k+i is based on the current state x k+i-1 , the control input u k+i-1 and the predicted demand for calculation.

[0197] Continuing, the goal of the MPC controller is to minimize the operating cost and meet the grid demand. The objective function (in quadratic form) can be expressed as:

[0198]

[0199] where the vector set of all control inputs within the prediction time domain is expressed as: x i -x refdenotes the state deviation at time i, i.e., the gap between the current state and the desired state; (x i -x ref ) T Q(x i -x ref ) represents the weighted quadratic norm of the state deviation, i.e., the gap between the current state and the desired state; Q represents the state deviation weighting matrix, usually a diagonal matrix, used to weight the deviations of each state; u i denotes the control input vector at time i; u i T Ru i represents the weighted quadratic norm of the control input, i.e., the magnitude of the control input; R represents the control input weighting matrix, usually a diagonal matrix, used to weight the magnitudes of each control input (Q and R are determined through empirical data or the tuning process). Meanwhile, to ensure the safety and stability of the system, the control input u i and the state x i need to satisfy certain constraint conditions (in linear form):

[0200]

[0201] Continuing, in the current step, this optimization framework uses a genetic algorithm to optimize the weighting matrices Q and R of the MPC in order to achieve optimal control within the prediction horizon. The genetic algorithm needs to go through the following specific steps: initialize the population, fitness evaluation, selection operation, crossover operation, mutation operation, generate a new population, until convergence.

[0202] 1. Initialize the population: Generate a number of individuals according to empirical data, and each individual represents a set of MPC parameters {Q, R}.

[0203] 2. Fitness function: Calculate the fitness value of each individual, i.e., the control performance under the given MPC parameters. The fitness function is defined as f(X * ) = -J, where the optimal control input sequence

[0204] 3. Selection: Select better individuals for reproduction according to the fitness values. The solution of the present invention adopts the roulette wheel selection method, that is, individuals with higher fitness have a higher probability of being selected.

[0205] First, calculate the total fitness based on the fitness value f(X * ) of each individual Then calculate the selection probability of each individual Finally, map the selection probability of each individual to the roulette wheel, and select individuals to form a new breeding population by generating random numbers between [0, 1) multiple times and according to the positions of the random numbers on the roulette wheel.

[0206] 4. Crossover: Perform crossover operations on some genes of the selected individuals to generate new individuals, so as to increase the diversity of the population and combine the excellent characteristics of the parent individuals.

[0207] For example: Q new = αQ1 + (1 - α)Q2, R new = βR1 + (1 - β)R2, where α and β are crossover coefficients.

[0208] 5. Mutation: Perform small-range random mutations on the new individuals generated by crossover to increase diversity and prevent the algorithm from falling into local optima. For example: Q mut = Q new + μQ, R mut = R new + μR, where μQ and μR are random perturbation momenta.

[0209] 6. Iteration: Repeat the operations of item selection, crossover, and mutation several times to gradually optimize the population until the fitness value no longer changes significantly. In the actual application scenario of the energy storage system, according to the specific requirements of the energy storage system, the calculation time and optimization effect can be balanced. If necessary, a relatively small maximum number of iterations can be set to achieve the real-time optimization effect of the architecture.

[0210] After the iteration ends, select the individual {Q * , R *} with the highest fitness as the optimal solution, and then combine the objective function and constraints, and solve the MPC optimization problem by using a quadratic programming (QP) solver to obtain the optimal control input sequence within the prediction horizon. Then apply the obtained optimal control input sequence to the energy storage system according to the PPS time reference to update the system state. Realize the minimization of the state deviation and control input cost of the system within the prediction horizon. Finally, the MPC controller continues to iterate and update, collect new state data x k+1 , and reconstruct the QP problem according to the new prediction requirements.

[0211] 4. Implementation of the control strategy:

[0212] Through the above process, under the unified PPS time reference, the optimized control strategy operation is applied to the energy storage system efficiently and accurately, realizing real-time adjustment of the system operation state and monitoring of the system operation effect, and periodically adjusting the prediction model and control strategy according to the feedback control strategy and market changes, meeting the interest balance among vehicle owners, energy storage operators, and the power grid.

[0213] ​In summary, in the V2G application scenario, as a buffer between EVs and renewable energy sources and the power grid, how to ensure the balance of interests among vehicle owners, energy storage operators, and the power grid, and consider the randomness of vehicle owners' charging behaviors, is a realistic problem currently faced by energy storage systems.

[0214] The technical solution of the present invention applies the time synchronization second pulse technology to the energy storage system, ensuring high synchronization inside and outside the system, improving the overall operation efficiency, and integrating the ARIMA and LSTM prediction models with a unified time reference. It can not only capture the linear trend of data but also handle complex non-linear characteristics, improving the accuracy of load prediction.

[0215] Then, by combining the MPC and genetic algorithm based on the prediction results, the on-time optimal control of the system is realized, improving the effectiveness and stability of the control strategy, ensuring that the control signal can be delivered and executed at the correct moment, and thus achieving higher prediction accuracy and adaptive control of the energy storage system.

[0216] At the same time, the second pulse technology ensures that all devices operate within the same time frame in complex application scenarios with a large number of devices, enabling the coordination and synchronization of the overall power system.

[0217] Compared with the prior art, this solution significantly improves the operation efficiency and stability of the energy storage system. By optimizing the control strategy, it reduces the overcharge and overdischarge behaviors of the energy storage system battery, extends the battery life, and optimizes the energy management strategy, reducing energy loss and operating costs, improving economic benefits, and thus achieving better performance in a dynamically changing market environment and participants' behaviors, providing more resilient support for the power grid.

[0218] The present invention has been described in detail above in combination with the embodiments with reference to the drawings. Those of ordinary skill in the art can make various variations to the present invention according to the above description. Therefore, certain details in the embodiments should not constitute a limitation to the present invention, and the protection scope of the present invention will be defined by the scope of the appended claims.

Claims

1. An energy storage system for V2G application scenarios, characterized in that: This energy storage system is based on time - synchronized second pulses, integrating a combined prediction model of two time - series prediction algorithms, ARIMA and LSTM, and equipped with a two - layer optimization architecture of an MPC controller and a genetic algorithm, thus forming a multi - layer composite optimization control strategy. This is used to enable the vehicle owner to, in the case of random charging behavior, achieve the energy buffering among the EV, the energy storage system, and the power grid according to the real - time data of the power grid to suppress supply - demand fluctuations and spatio - temporal mismatches. The time - synchronized second pulse uses a high - precision timing reference to ensure that data acquisition and control operations of each node of the energy storage system are carried out under the same time reference, ensuring the timing consistency of data acquisition, processing, and control decisions of the entire system. The system hardware of the time - synchronized second pulse includes: a BD / GPS dual - mode receiver, a high - precision clock synchronization module, and a disciplined clock sub - module. The BD / GPS dual - mode receiver receives the standard time signal transmitted by satellites and generates a 1PPS signal through the processing of the high - precision clock synchronization module. The BD / GPS dual - mode receiver is connected to and deployed in the energy storage system. The time - synchronization interface of the high - precision clock synchronization module outputs the 1PPS signal to the corresponding interface of the energy control main board of the energy storage system to ensure that all control modules of the energy storage system maintain clock synchronization under a unified and high - precision time reference. The disciplined clock sub - module is deployed in the V2G bi - directional charging pile, connected to its control module, and connected to the time - synchronization interface of the high - precision clock synchronization module of the energy storage system. It is calibrated and disciplined by receiving the 1PPS signal provided by it to achieve clock synchronization consistent with the energy storage system. The ARIMA prediction model is used to analyze and predict linear time - series data in a time series, and identify and capture the trends and seasonal variations of the data through the historical charging data of electric vehicles. The LSTM prediction model is used to learn and predict time - series data with non - linear and long - term dependence relationships in a time series, and can effectively capture the randomness of the vehicle owner's charging behavior, including changes with factors such as time, temperature, and working day / holiday. The combined prediction model of the two prediction algorithms can provide a more accurate charging load prediction for the energy storage system, making the energy storage system more intelligent in the face of random and volatile demand changes at different time scales. It can synchronize the system operation with the power grid load fluctuations, accurately optimize the charging and discharging strategies, use the ARIMA model to process short - term linear dependencies, and use the LSTM model to capture long - term dependencies and non - linear patterns such as complex patterns. The role of the time - synchronized second pulse in this combined prediction model is to ensure that the time stamps of all time - series data are synchronized with the positions when actual events occur, thereby improving the accuracy and reliability of the prediction. In the two - layer optimization architecture of the MPC controller and the genetic algorithm, where: The MPC controller generates an optimal charging / discharging strategy in real - time according to the results provided by the combined prediction model of ARIMA and LSTM to minimize the operating cost and meet the power grid demand. The genetic algorithm is used to optimize the parameters of the MPC controller to ensure finding the optimal control strategy in the V2G environment, that is: by taking the control strategy generated by the MPC controller as the initial population of the genetic algorithm, randomly generating an initial parameter set, and calculating the appropriate fitness function for each individual, evaluating the economic benefits of the vehicle owner, energy storage operator, and power grid, and performing selection, crossover, and mutation genetic operations on individuals with high fitness to generate a new generation of population, that is, the optimal control strategy that satisfies all of the EV, energy storage system, and power grid.

2. A energy storage system for V2G application scenarios according to claim 1, characterized in that: The energy storage system receives a time synchronization signal at the beginning of each whole second, compares and calibrates this signal with the clocks of each control module inside the system to ensure that all control modules in the system operate on the same time standard. Within each synchronization cycle, the high-precision clock synchronization module 1PPS updates the clocks of each tamed clock sub-module for time calibration and deviation correction. The specific algorithm is as follows: ΔT i = T BD / GPS - T local,i Where: T BD / GPS represents the standard time of BD / GPS received by the high-precision clock synchronization module, i.e., 1PPS, T local,i is the local clock time of the i-th subsystem, ΔT i is the measured time deviation, is the corrected local time.

3. An optimized control strategy for an energy storage system according to claim 1, comprising the following steps: Step 1, receive the BD / GPS time synchronization signal for second pulse time synchronization; Step 2, collect charge and discharge time series data; Step 3, training of the combined prediction model. Use the ARIMA model to process the short-term linear relationship of the charging behavior data, and then use these prediction results as input and combine other features, including temperature and date type, and use the LSTM model to capture and predict long-term, non-linear patterns; Step 4, real-time load prediction. Gradually reduce the prediction error by iteratively updating the weights and biases to achieve a more accurate prediction of future demand; Step 5, generation of the MPC control strategy. The MPC controller generates an optimal charge / discharge control strategy for the energy storage system based on the prediction results of the time series relationship generated by the ARIMA and LSTM models, and can be adjusted according to real-time data to ensure the stable operation and optimal control of the system in a dynamic environment; Step 6, genetic algorithm optimization. The genetic algorithm optimizes the parameters of the MPC controller to avoid falling into local optima. Through the iterative evolution of the genetic algorithm, achieve adaptive adjustment of the control strategy and improve the control effect; Step 7, generation of the charge / discharge strategy and execution of the charge / discharge strategy.

4. An optimized control strategy for an energy storage system according to claim 3, wherein the ARIMA model in step 3 is specifically as follows: Short-term linear prediction using the ARIMA model: The ARIMA model consists of three parameters p, d, and q: p is the autoregressive order; d is the differencing order; q is the moving average order. Determine the parameters d, p, and q of the ARIMA model through stationarity tests, calculations, and observation of the autocorrelation function and partial autocorrelation function graphs; d is determined by the stationarity test, the ADF test, that is, use the ADF test to determine the stationarity of the time series; ADF test formula: Δy t = α + βt + γy t-1 + δ1Δy t-1 + δ2Δy t-2 +…+ δ k Δy t-k + ε t where, Δy t is the difference term, α is a constant, βt is the time trend term, γy t-1 is the lag term, δ i Δy t-i is the lag difference term, ε t is the random error term The autoregressive order p and the moving average order q are determined by observing the autocorrelation function ACF and the partial autocorrelation function PACF. ACF formula: where: ρ k is the autocorrelation coefficient at lag k; y t is the value of the time series at time t; y t-k is the value of the time series at time t - k, representing the value k time points back from time t, is the mean of the time series; T is the total length of the time series; PACF formula: Where: φ kk is the partial autocorrelation coefficient at lag k, representing the autocorrelation of the sequence excluding the intermediate lags at lag k; ρ k is the autocorrelation coefficient at lag k, representing the linear correlation between the value of the sequence at the current moment and the value at lag k moments; φ k-1,i is the partial autocorrelation coefficient at lag k-1, which is an intermediate term to be considered when calculating the partial autocorrelation coefficient at the current lag k; ρ k-i is the autocorrelation coefficient at lag k-i, used to adjust the value of the partial autocorrelation coefficient at the current lag k, ARIMA Model Construction and Prediction: Input the estimated parameter values into the ARIMA model to establish an ARIMA(p,d,q) model. Through the combined action of differencing, regression, and moving average, short-term linear prediction of time series data can be performed, which can be used for predicting charging demand.

5. An optimized control strategy for an energy storage system according to claim 3 or 4, wherein the LSTM model in step 3 is specifically as follows: Using LSTM for Long-Term Nonlinear Prediction: The LSTM model controls the flow of information through three "gate" structures, namely: the forget gate, the input gate, and the output gate, so as to effectively preserve long-term dependencies; First, the forget gate f t , which is used to determine what information to discard from the cell state f t = σ(W f ·[h t-1 , X t + b f ) Wherein: W f is the weight matrix of the forget gate; [h t-1 , X t is the connection between the hidden state at the previous time point and the current input; b f is the bias of the forget gate; σ is the sigmoid activation function, which outputs a scalar value between 0 and 1, representing the retention ratio; Second, the input gate i t and the candidate cell state i t = σ(W i · [h t-1 , X t + b i ),Function: Determine what new information to store in the cell state; where W i , W c is the weight matrix of the input gate; b i , b c is the bias term of the input gate; tanh is the hyperbolic tangent activation function, which is used to generate the candidate cell state, and the output range is between -1 and 1; Third, cell state update Y t , for updating the cell state, combining forgotten and newly input information where Y t is the cell state at the current time point; Y t-1 is the cell state at the previous time point; Fourth, output gate o t and hidden state h t : o t = σ(W o ·[h t-1 , X t + b o ), which is used to determine what information to output at this time point. h t = o t * tanh(Y t ), that is, finally output the actual demand prediction through the transfer function Where: W o is the weight matrix of the output gate; b o is the bias term of the output gate.

6. An optimized control strategy for an energy storage system according to claim 5, wherein the calculation steps of the LSTM model in step 3 are as follows: Step A, Initial Step: Features related to charging requirements include: Create the feature set X based on temperature, time, and holidays t , determine the hidden layer, which contains a number of LSTM units, and each unit captures the temporal dependencies in the data Step B, Data Processing: The prediction result C of the ARIMA model t+1 is combined with other features X t , including data from t-n to t, to construct the training set [X t-n ,…,X t ,C t+1 ; Step C, LSTM Model Training: The training of the LSTM model is an iterative process. It uses historical time series data to train the LSTM network through steps such as forward propagation, loss function, backpropagation, and parameter update to determine the weight matrix and bias values; Among them, the goal of the loss function L is to minimize the error between the predicted output and the actual demand, and the mean square error MSE is adopted: Where: Z t+1 is the actual charging demand value, is a predicted charging demand value output by LSTM at each step of iteration based on the current weights and biases using the training set, and N is the number of samples. Backpropagation is used to calculate the gradient of the loss function with respect to each model parameter (weights and biases). The gradient represents the rate of increase of the loss function near the current value of the parameter and is used to update the weights to reduce the error. Parameter update is calculated using the gradient descent method: where α is the learning rate and L is the loss function, and are the gradients of the weight matrix and the bias respectively, where W old and b old are the unified expressions of the weight matrices and bias values of the three "gates" respectively, and are the updated weight matrix and bias value respectively, which are updated and adjusted according to the gradient of the loss function in each training iteration, so that the model can better learn and capture the underlying patterns of the data; Step D, Prediction: The training set [X t-n , …, X t , C t+1 input into the LSTM model to the LSTM formula after updating the weight matrix and bias value and calculate the final prediction at time t+1 through the conversion function 7. An optimized control strategy for an energy storage system according to claim 5, wherein the MPC controller in step 5 is controlled through the following process: First, the dynamic behavior of the energy storage system can be described by a linear state space model: The output equation is: y k = Dx k + v k Among them, x k is the system state vector at time k; u k is the control input vector at time k; is the predicted electricity demand at time k+1 predicted by the ARIMA and LSTM models; A is the state transition matrix; B is the control matrix; C is the input matrix; y k is the system output vector at time k; w k is the process noise; D is the output matrix; v k is the measurement noise; Second, the goal of MPC is to optimize the control input u within the prediction time domain predicted by the ARIMA and LSTM models k , such that the system state x k tracks the desired state x ref ; Given that the prediction time domain length is N, the control models for the predicted states and control inputs at the next N time instants can be expressed as: Among them is the sequence of predicted state vectors from time k + 1 to k + N; i = 1, 2, 3…N, is the predicted electricity demand at time k + i, which is predicted by the ARIMA and LSTM models; The model calculates future states through recursion, where each predicted state x k+i is based on the current state x k+i-1 , the control input u k+i-1 and the predicted demand and is calculated accordingly. Third, the goal of the MPC controller is to minimize the operating cost and meet the grid demand, and the objective function is expressed as: Among them, the vector set of all control inputs within the prediction horizon is expressed as: x i -x ref represents the state deviation at time i, that is, the gap between the current state and the desired state; (x i -x ref ) T Q(x i -x ref ) represents the weighted quadratic norm of the state deviation, that is, the gap between the current state and the desired state; Q represents the state deviation weighting matrix; u i represents the control input vector at time i; represents the weighted quadratic norm of the control input, that is, the magnitude of the control input; R represents the control input weighting matrix; the control input u i and the state x i need to satisfy certain constraint conditions:

Images