Energy storage power station temperature early warning method based on LSTM and SVM

Abstract

The invention discloses an energy storage power station temperature early warning method based on LSTM and SVM. The method comprises the following steps: selecting characteristic parameters influencing the battery temperature as the input of an LSTM and SVM battery temperature hybrid prediction model; respectively constructing battery temperature prediction models based on LSTM and SVM; according to the historical data, the LSTM and SVM battery temperature prediction models are adopted to carry out temperature prediction; fusing the prediction results of the LSTM and the SVM by using a particle swarm optimization algorithm to obtain a comprehensive prediction result of the battery temperature; and according to the battery temperature comprehensive prediction result, graded early warning actions are adopted. According to the method, the battery state can be accurately predicted, a technician can eliminate possible faults in advance, the accident rate is reduced, and the safety of the energy storage system is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of temperature early warning of energy storage power stations, and in particular to a temperature early warning method of energy storage power stations based on LSTM and SVM. Background Art

[0002] Environmental pollution and energy shortages are becoming increasingly serious. The scale of green energy construction is increasing. However, the randomness, intermittency and imbalance of renewable energy generation represented by wind power generation and photovoltaic power generation have led to large-scale new energy grid connection, which poses a huge challenge to the safety of the power grid. Against this background, electrochemical energy storage came into being. Electrochemical energy storage is a key technology to promote the emergence of new energy, and it is also an important method to improve the flexibility, economy and safety of traditional power grids. With the continuous increase in the installed capacity of electrochemical energy storage power stations, there are more and more safety issues in construction. The series-parallel structure of the energy storage power station is simple, but the overall capacity and power are large, and there are many safety risk points. Therefore, the safety research of energy storage power stations is the current key point and difficulty.

[0003] The problem of energy storage battery runaway has not been completely solved in practical applications. When energy storage batteries are subjected to mechanical abuse, electrical abuse, and thermal abuse, the battery materials are damaged and abnormal heating occurs. The continuous accumulation of heat intensifies the internal exothermic chemical reaction, forming a positive feedback, and ultimately causing thermal runaway. A large amount of heat and flammable gas is released during the thermal runaway process. When the density of the flammable gas reaches its explosion limit, it will explode under the action of external high temperature, seriously affecting the safe and stable operation of the energy storage power station. Therefore, it is urgent to study the thermal runaway characteristics of energy storage batteries, explore early warning methods for thermal runaway, and solve the problem of safety management and control of energy storage power stations.

[0004] With the development of intelligent technology, deep learning algorithms are introduced into energy storage power stations. By building a temperature prediction model for energy storage batteries through deep learning algorithms, abnormal battery temperature operation can be discovered in advance, improving the safety and reliability of energy storage power station grid connection. There are many statistical and deep learning algorithms that can be used for battery temperature prediction, such as Bayesian models, neural networks, decision trees, etc. However, the prediction performance of these traditional single prediction algorithms is not stable, the accuracy of battery temperature prediction cannot be well maintained, and the effective time of the prediction model is relatively short. Summary of the invention

[0005] In order to overcome the above-mentioned deficiencies of the prior art, the present invention provides a temperature warning method for an energy storage power station based on LSTM and SVM, which is used to solve some deficiencies of the existing energy storage power station for early warning of battery temperature and improve the accuracy of battery temperature prediction.

[0006] The technical solution of the present invention to solve the above technical problems is as follows: a temperature early warning method for an energy storage power station based on LSTM and SVM, comprising the following steps:

[0007] S1. Select the battery external characteristic parameters that affect the temperature of the single battery as the input of the LSTM and SVM battery temperature hybrid prediction model;

[0008] S2, build a battery temperature prediction model based on LSTM and SVM;

[0009] S3, based on the historical voltage, SOC and temperature data of the single battery, the LSTM model is used to predict the future value of the battery temperature;

[0010] S4. Perform linear and nonlinear analysis based on historical data and use the SVM model to predict the battery temperature.

[0011] S5. Use the particle swarm optimization algorithm to fuse the prediction results of LSTM and SVM to obtain a comprehensive prediction result of the battery temperature;

[0012] S6. Take graded warning actions based on the comprehensive prediction results of battery temperature.

[0013] Preferably, the present invention selects historical single cell voltage U, temperature T, SOC as battery external characteristic parameters affecting single cell temperature and inputs of the hybrid prediction model.

[0014] Furthermore, step S2 includes the following sub-steps:

[0015] A1. Sort the data to obtain the sorted parameter time series;

[0016] A2. Normalize the data;

[0017] A3. Divide the data into training data set and test data set;

[0018] A4. Use the training data set to train the LSTM and SVM prediction models to obtain the battery temperature prediction models based on LSTM and SVM respectively.

[0019] Furthermore, the step S4 comprises:

[0020] Input historical data, use regression analysis to make linear predictions on battery temperature, compare the residual value with the actual value, and use SVM to make nonlinear predictions. The sum of the two is the overall prediction of battery temperature.

[0021] Furthermore, the step S5 comprises:

[0022] B1. The objective function and constraints of establishing the LSTM and SVM hybrid prediction model are:

[0023]

[0024] y r is the actual battery temperature value; is the output of the LSTM and SVM prediction models; k1, k2 are the weights of a single model.

[0025] B2. Randomly assign an initial position and velocity to each particle, and find the optimal solution for the two distribution coefficients:

[0026]

[0027]

[0028] After calculating f1 and f2, compare the size relationship between the two and take the position corresponding to the smaller value as the current global optimal solution.

[0029] gbest=argmin{f1,f2}

[0030] pbest is the local optimal solution, and gbest is the global optimal solution.

[0031] B3, Update the speed and position of the particle. Substitute the gbest and pbest obtained in B2 into the following formula to update the particle:

[0032] x i,new =x i,old +v i,new

[0033] v i,new =ω×v i,old +c1×r×(pbest i -x i )+c2×r×(gbest i -x i )

[0034] c1, c2 are learning coefficients, r is a random number with a value of [0,1], x i is the current position of the particle.

[0035] B4. Substitute the updated particle position into the objective function to obtain the new f1 * , Compare their sizes, and the smaller value is taken as the current global optimal solution:

[0036]

[0037] B5. The optimal weight distribution is k1, k2. At this time, the prediction result of the hybrid model is:

[0038]

[0039] The step S6 comprises:

[0040] C1. The upper limit of the single-cell charging temperature is 55°C, the upper limit of the single-cell discharging temperature is 55°C, the lower limit of the single-cell charging temperature is 5°C, the lower limit of the single-cell discharging temperature is -5°C, and the temperature change rate per unit time does not exceed 0.5;

[0041] C2. When the predicted battery temperature is within the set threshold, the temperature change rate is calculated. If the temperature change rate exceeds the set threshold, the system will issue an early warning, otherwise no early warning action will be taken;

[0042] C3. When the predicted battery temperature is outside the set threshold, the system sends a warning signal.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. Based on neural network and support vector machine algorithms, the present invention considers the characteristic parameters that affect the battery temperature change, constructs a battery temperature prediction model based on LSTM from the perspective of time series and associated time series, and analyzes from linear and nonlinear perspectives respectively, constructs a battery temperature prediction model based on SVM, and uses the particle swarm optimization algorithm to calculate the optimal distribution coefficient of the two models, which can realize the accurate prediction of battery temperature to the greatest extent and reduce the probability of thermal runaway of energy storage batteries.

[0045] 2. The present invention sets a suitable threshold for the temperature of the energy storage battery and performs corresponding alarm actions on the battery module with temperature exceeding the limit according to the temperature prediction value obtained by the model prediction, which effectively simplifies the workload of maintenance personnel and reduces the cost of failures, thereby improving the economy and safety of the energy storage power station. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a flow chart of the present invention.

[0047] Figure 2 LSTM network model diagram for battery temperature prediction of the present invention.

[0048] Figure 3 This is a framework diagram of the SVM battery temperature prediction provided by the present invention.

[0049] Figure 4 This is a basic framework diagram of the hybrid prediction model provided by the present invention. Specific implementation methods

[0051] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0052] See also Figure 1 , Figure 1 The specific implementation steps of the present invention are as follows:

[0053] S1. Select the battery external characteristic parameters that affect the temperature of the single cell as the input of the LSTM and SVM battery temperature hybrid prediction model; this embodiment selects the single cell in the energy storage system as the research object, and the collected signals include the single cell voltage, temperature, and SOC as the input of the prediction model.

[0054] S2. Construct the battery temperature prediction models of LSTM and SVM respectively.

[0055] The process of building a battery temperature prediction model is as follows:

[0056] A1. Collect historical data of single cell voltage, temperature, and SOC, sort the data, and obtain the sorted time series of each parameter;

[0057] A2. Normalize the data;

[0058] A3. Divide the normalized data set into a training set and a test set with a ratio of 7:3;

[0059] A4. Use the training set to train the LSTM battery temperature prediction model, and use the test set to test it until the expected effect is achieved. When the error of the battery temperature T is less than 0.01, output the trained prediction model; (the previous description is to train both the LSTM and SVM prediction models)

[0060] Figure 2 middle is the prediction result output by LSTM, I in It is the input of the LSTM model, which consists of historical battery voltage, temperature, and SOC.

[0061] According to the influencing parameters that currently affect the battery temperature change, they are input into the LSTM temperature prediction model established above, and the current predicted value of the battery temperature is output.

[0062] The LSTM network model performs parameter prediction on battery temperature. It predicts future battery temperature values ​​by inputting historical single-cell battery voltage, temperature, and SOC. By inputting other variables that affect battery temperature changes, it not only analyzes the temperature time series, but also considers the impact of different influencing factors on battery temperature, thereby improving the accuracy of the prediction.

[0063] S3, predicting the future value of the battery temperature using the LSTM battery temperature prediction model constructed according to the above steps;

[0064] S4. Perform linear and nonlinear analysis based on historical data and use the SVM model to predict the battery temperature.

[0065] The present invention uses regression analysis combined with SVM to predict the battery temperature. The law of battery temperature change can be decomposed into linear and nonlinear parts. Linear regression analysis is responsible for predicting linear problems in temperature change. By comparing the true value, the residual part of the temperature prediction at time t can be obtained. This residual part is nonlinear. The residual value is used as the input of the SVM prediction model for fitting prediction, and the two parts of the prediction results are superimposed to obtain the final battery temperature prediction value.

[0066] S5. Use the particle swarm optimization algorithm to fuse the prediction results of LSTM and SVM to obtain a comprehensive prediction result of the battery temperature.

[0067] Since a single prediction model has its own limitations, it is difficult to achieve satisfactory results. Therefore, the present invention proposes a hybrid prediction model, which uses the particle swarm optimization (PSO) algorithm to assign optimal weights to the two models, so that the LSTM and SVM prediction models are better integrated, the overall prediction accuracy of the model is improved, and the effectiveness and reliability of the model are increased.

[0068] like Figure 3 As shown, is the linear prediction result, is the residual part captured by the LR model, x1 and x2 are variables that affect the battery temperature, θ is the regression equation parameter, ε is the error, g is the nonlinear model based on SVM, and Δt is the random error. The prediction of battery temperature is divided into linear and nonlinear part predictions. Linear regression analysis is responsible for analyzing linear problems in the prediction, finding the linear relationship between battery temperature and voltage and SOC, and realizing linear prediction of battery temperature. By comparing the true value, the residual part of the temperature prediction at time t can be obtained. This residual part is nonlinear. The residual value is used as the input of the SVM model for fitting prediction. The linear prediction result of the final regression analysis plus the nonlinear prediction result of the SVM obtains the final battery temperature prediction value.

[0069] The LSTM-SVM hybrid prediction model framework based on particle swarm algorithm is as follows Figure 4 The particle swarm algorithm is used to better integrate the prediction results of the two models, avoiding the defects brought by a single model.

[0070] B1. The objective function and constraints of establishing the LSTM and SVM hybrid prediction model are:

[0071]

[0072] y r is the actual battery temperature value; is the output of the LSTM and SVM prediction models; k1 and k2 are the weights of a single model.

[0073] B2. Randomly assign an initial position and velocity to each particle, and find the optimal solution for the two distribution coefficients:

[0074]

[0075]

[0076] After calculating f1 and f2, compare the size relationship between the two and take the position corresponding to the smaller value as the current global optimal solution.

[0077] gbest=argmin{f1,f2}

[0078] pbest is the local optimal solution, and gbest is the global optimal solution.

[0079] B3, Update the speed and position of the particle. Substitute the gbest and pbest obtained in B2 into the following formula to update the particle:

[0080] x i,new =x i,old +v i,new

[0081] v i,new =ω×v i,old +c1×r×(pbest i -x i )+c2×r×(gbest i -x i )

[0082] c1, c2 are learning coefficients, r is a random number with a value of [0,1], x i is the current position of the particle.

[0083] B4. Substitute the updated particle position into the objective function to obtain the new f1 * , Compare their sizes, and the smaller value is taken as the current global optimal solution:

[0084]

[0085] B5. The optimal weight distribution is k1, k2. At this time, the prediction result of the hybrid model is:

[0086]

[0087] S6. Take graded warning actions based on the comprehensive prediction results of battery temperature.

[0088] After the final battery temperature prediction result is calculated by the above method, the corresponding alarm action is performed in combination with the set temperature threshold. The steps are as follows:

[0089] C1. The upper limit of the single-cell charging temperature is 55°C, the upper limit of the single-cell discharging temperature is 55°C, the lower limit of the single-cell charging temperature is 5°C, the lower limit of the single-cell discharging temperature is -5°C, and the temperature change rate per unit time does not exceed 0.5;

[0090] C2. When the predicted battery temperature is within the set threshold, the temperature change rate is calculated. If the temperature change rate exceeds the set threshold, the system will issue an early warning, otherwise no early warning action will be taken;

[0091] C3. When the predicted battery temperature is outside the set threshold, the system sends a warning signal.

Claims

1. A temperature early warning method for energy storage power station based on LSTM and SVM, characterized in that: The following steps are involved: S1. Select the battery external characteristic parameters that affect the temperature of the single battery as the input of the LSTM and SVM battery temperature hybrid prediction model; S2, build a battery temperature prediction model based on LSTM and SVM; S3, based on the historical voltage, SOC and temperature data of the single battery, the LSTM model is used to predict the future value of the battery temperature; S4. Perform linear and nonlinear analysis based on historical data and use the SVM model to predict the battery temperature. S5. Use the particle swarm optimization algorithm to fuse the prediction results of LSTM and SVM to obtain a comprehensive prediction result of the battery temperature; S6. Take graded warning actions based on the comprehensive prediction results of battery temperature.

2. The temperature early warning method for energy storage power station based on LSTM and SVM as claimed in claim 1 is characterized in that: The external characteristic parameters that serve as the input of the LSTM and SVM prediction models and affect the battery temperature include: historical single cell voltage U, temperature T, and SOC.

3. The temperature early warning method for energy storage power station based on LSTM and SVM as claimed in claim 1, characterized in that: The step S2 comprises: A1. Sort the data to obtain the sorted parameter time series; A2. Normalize the data; A3. Divide the data into training data set and test data set; A4. Use the training data set to train the LSTM and SVM prediction models to obtain the battery temperature prediction models based on LSTM and SVM respectively.

4. The temperature early warning method for energy storage power station based on LSTM and SVM as claimed in claim 1 is characterized in that: In step S4, historical data is input, and linear prediction of battery temperature is performed using regression analysis. The residual value is compared with the actual value and nonlinear prediction is performed using SVM. The sum of the two is the overall prediction of battery temperature.

5. The temperature early warning method for energy storage power station based on LSTM and SVM as claimed in claim 1, characterized in that: Use the particle swarm optimization algorithm to assign the most effective weights to the prediction results output by the LSTM model and the SVM model, and output more accurate prediction results. The specific steps are as follows: B1. The objective function and constraints of establishing the LSTM and SVM hybrid prediction model are: y r is the actual battery temperature value; is the output of LSTM and SVM prediction models; k1, k2 are the weights of a single model; B2. Randomly assign an initial position and velocity to each particle, and find the optimal solution for the two distribution coefficients: After obtaining f1 and f2, compare the size relationship between the two and take the position corresponding to the smaller value as the current global optimal solution. gbest=argmin{f1,f2} pbest is the local optimal solution, gbest is the global optimal solution; B3. Update the speed and position of the particle. Substitute the gbest and pbest obtained in B2 into the following formula to update the particle: x i,new =x i,old +v i,new v i,new =ω×v i,old +c1×r×(pbest i -x i )+c2×r×(gbest i -x i ) c1, c2 are learning coefficients, r is a random number with a value of [0,1], x i is the current position of the particle; B4. Substitute the updated particle position into the objective function to obtain the new Compare their sizes, and the smaller value is taken as the current global optimal solution: B5. The optimal weight distribution is k1, k2. At this time, the prediction result of the hybrid model is:

6. The temperature early warning method for energy storage power station based on LSTM and SVM as claimed in claim 1, characterized in that: The step S6 is specifically as follows: C1. The upper limit of the single-cell charging temperature is 55°C, the upper limit of the single-cell discharging temperature is 55°C, the lower limit of the single-cell charging temperature is 5°C, the lower limit of the single-cell discharging temperature is -5°C, and the temperature change rate per unit time does not exceed 0.5; C2. When the predicted battery temperature is within the set threshold, the temperature change rate is calculated. If the temperature change rate exceeds the set threshold, the system will issue an early warning, otherwise no early warning action will be taken; C3. When the predicted battery temperature is outside the set threshold, the system sends a warning signal.

Images