Energy storage system control method based on multi-scale feature fusion and dynamic adjustment

Abstract

The invention relates to an energy storage system control method based on multi-scale feature fusion and dynamic adjustment, and belongs to the technical field of energy storage regulation and control. The method comprises the steps that historical new energy power generation power and power load data are acquired, a net load sequence is generated, and zero-mean standardization is carried out; the number of wavelet decomposition layers is optimized based on mutual information entropy, and the standardized net load sequence is adaptively divided into a high-frequency component set and a low-frequency component set; predicting a high-frequency component by adopting a weighted moving average method, and predicting a low-frequency component by adopting linear regression; and finally, fusing the prediction results of the high-frequency and low-frequency components, and calculating the output value of the energy storage system at the next moment. Through data preprocessing, wavelet decomposition layer number optimization and multi-scale prediction fusion, dynamic and accurate adjustment of the output of the energy storage system is realized, and the adaptability of the energy storage system to new energy fluctuation and load change is improved.

Description

Technical Field

[0001] This invention belongs to the field of energy storage regulation technology, and in particular to a control method for energy storage systems based on multi-scale feature fusion and dynamic adjustment. Background Technology

[0002] With the accelerated pace of the global energy transition, the penetration rate of renewable energy sources such as wind and solar power in the power system is rapidly increasing. However, the large-scale integration of this green energy has also brought new technological challenges: new energy power generation exhibits significant intermittent, random, and volatile characteristics, with output changes spanning multiple time scales, from instantaneous fluctuations at the second and minute levels to regular changes at the hour and day levels, and even long-term seasonal trends. This multi-time-scale characteristic makes the net load sequence of the power system exceptionally complex, making it difficult for traditional prediction and control methods based on a single time scale to effectively address the issue. Specifically, high-frequency fluctuations may be caused by sudden weather changes (such as cloud cover and gusts), while low-frequency trends are related to macroeconomic factors such as seasonal changes and climate patterns. As a key regulatory means to balance power generation and consumption and maintain the stable operation of the power grid, the accuracy of the control strategy of energy storage systems directly affects the safety and economy of the entire power system. Current widely used energy storage control methods, such as charge-discharge strategies based on fixed threshold rules or simple moving average prediction models, often fail to adequately adapt to the multi-scale, non-stationary load characteristics, easily leading to problems such as control lag or over-regulation. Therefore, there is an urgent need to develop a novel energy storage control method capable of intelligently identifying and integrating multi-scale characteristics.

[0003] The current technical solutions in the field of power system energy storage control can be divided into the following categories: (1) Traditional time series forecasting methods, such as the autoregressive integral moving average (ARIMA) model and exponential smoothing method, have the main advantages of simple model structure, high computational efficiency, and easy engineering implementation. However, these methods are essentially based on linear assumptions and have limited ability to model new energy output sequences that exhibit strong nonlinearity and non-stationary characteristics. In particular, they are difficult to effectively capture sudden fluctuations and multi-time scale characteristics. (2) Rule-based control strategies, such as fixed threshold charge and discharge control, are simple to implement and have a fast response, but lack a forward-looking judgment of the system's operating state. Under complex operating conditions, they are prone to poor control effects or frequent equipment operation. (3) With economic or technical objectives, intelligent algorithms are used to solve the problem based on constraints, such as neural networks and support vector machines. Although they have strong nonlinear fitting capabilities, they usually require a large amount of training data and have problems such as poor model interpretability and complex parameter adjustment. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and propose a control method for energy storage systems based on multi-scale feature fusion and dynamic adjustment. After preprocessing historical new energy power generation and load data, a standardized net load sequence is obtained. Based on mutual information entropy, the wavelet decomposition level is optimized, and the net load sequence is adaptively divided into high-frequency and low-frequency components. Weighted moving average and linear regression are used for prediction respectively. Finally, the high-frequency and low-frequency prediction results are fused to dynamically adjust the output of the energy storage system.

[0005] The technical problem solved by this invention is achieved through the following technical solution: A control method for energy storage systems based on multi-scale feature fusion and dynamic adjustment includes the following steps: Step 1: Obtain historical renewable energy power generation and power load data, and obtain the net load sequence; Step 2: Data preprocessing and outlier correction, and zero-mean standardization of the net load series; Step 3: Optimize the number of layers in the net load sequence after zero-mean standardization based on wavelet decomposition using mutual information entropy; Step 4: Based on the obtained number of layers, divide the high-frequency component set and low-frequency component set after wavelet decomposition; Step 5: Predict high-frequency components using the weighted moving average method; Step 6: Use linear regression to predict low-frequency components; Step 7: Combine the prediction results of high-frequency and low-frequency components to calculate the output value of the energy storage system at the next moment.

[0006] Furthermore, the specific implementation method of step 1 is as follows: obtain the historical renewable energy power generation sequence. N ( t )( t =1,2,3..., T all , T all (Total length of historical data), power load data L ( t Then the net load sequence P ( t )for: 。

[0007] Furthermore, the specific implementation method of step 2 is as follows: The net load sequence is standardized to zero mean to eliminate the influence of dimensions; the net load sequence after zero mean standardization is... P norm ( t ): in,u P The arithmetic mean of the net load series. s P is the standard deviation of the net load series.

[0008] Furthermore, the specific implementation method of step 3 is as follows: select the db4 wavelet basis, and set the initial decomposition level. N =3, for standardized sequences P norm ( t Perform wavelet decomposition to obtain approximate components. A N Low-frequency trends and detail components D 1, D 2,..., D N ,satisfy P norm ( t )= A N + D 1+ D 2+...+ D N ; For adjacent detail components D k and D k+1 Calculate mutual information entropy E ( D k , D k+1 ): in, p ( x , y ) represents the joint probability density. p ( x ), p ( y ) represents the marginal probability density; if it exists E ( D k , D k+1 () greater than the mutual information entropy threshold e To indicate the degree of scale aliasing, then N = N Repeat the decomposition and calculation by adding 1 until all adjacent components are reached. E All are less than or equal to the mutual information entropy threshold. e This indicates that the optimal number of layers has been obtained. N * .

[0009] Furthermore, the specific implementation method of step 4 is as follows: based on the obtained number of layers, divide the high-frequency component set and the low-frequency component set after wavelet decomposition, wherein the high-frequency component set... H Include D 1 to D K ,in K less than or equal to N Integers of * / 2 are used to capture short-term fluctuation characteristics; low-frequency component set L Include D K+1 to D N* as well as A N* Used to capture long-term trend characteristics, ultimately yielding high-frequency components. H ( t )= D 1+ D 2+...+ D K Low-frequency components L ( t )= A N* + D K + D K+1 +...+ D N* .

[0010] Furthermore, step 5 employs a weighted moving average method to predict high-frequency components. H pred ( t The specific implementation method for +1 is as follows: in, w ti Indicates the first you The weights of historical data are determined by fitting the initial weight values ​​to the historical data, with more recent data having higher weights.

[0011] Furthermore, step 6 uses linear regression to predict the low-frequency components. L pred ( t The specific implementation method for +1 is as follows: in, i 0、 i 1 represents the coefficient of fit for linear regression.

[0012] Furthermore, step 7 integrates the prediction results of high-frequency and low-frequency components to calculate the output value of the energy storage system in the power system at the next moment. Pfinal ( t+ The specific implementation method of 1) is as follows: when H pred ( t +1)+ L pred ( t When +1)≥0: when H pred ( t +1)+ L pred ( t When +1) < 0: in, P rate This is the rated power of the energy storage system.

[0013] The advantages and positive effects of this invention are: 1. This invention automatically determines the number of decomposition layers of the wavelet decomposition algorithm through mutual information entropy, thereby achieving accurate separation of short-term fluctuations and long-term trends in net load power data, reducing scale aliasing errors, and demonstrating strong multi-scale feature decoupling capabilities.

[0014] 2. This invention predicts the components after algorithm decomposition separately, uses a weighted moving average method to capture sudden fluctuations for high-frequency components, and combines linear regression to extract trend features for low-frequency components. This avoids the shortcomings of a single model that cannot take into account both sudden changes and trends, and ensures that the output of the energy storage system can better meet the needs of various scenarios. Attached Figure Description

[0015] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0016] The present invention will be further described in detail below with reference to the accompanying drawings.

[0017] A control method for energy storage systems based on multi-scale feature fusion and dynamic adjustment, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain historical renewable energy power generation and power load data, and obtain the net load sequence.

[0018] Obtain historical renewable energy power generation sequence N ( t )( t =1,2,3..., T all , T all(Total length of historical data), power load data L ( t Then the net load sequence P ( t )for: Step 2: Data preprocessing and outlier correction, and zero-mean standardization of the net load series.

[0019] Zero-mean standardization of the net load series eliminates the influence of dimensions. The zero-mean standardized net load series is as follows: P norm ( t ): Among them, the arithmetic mean of the net load series u P Standard deviation of net load series s P Calculated using the following method: If | P norm ( t If the value is greater than 1.5, it is considered an outlier and outlier handling is performed (e.g., sudden changes caused by equipment failure or extreme weather). A sliding window linear interpolation method is used for correction. In the formula, m Take half the length of the window, such as m Take 3.

[0020] Step 3: Optimize the number of layers of the net load sequence after zero-mean standardization based on wavelet decomposition using mutual information entropy.

[0021] The optimal number of layers in adaptive wavelet decomposition is dynamically determined by the mutual information entropy of adjacent scale components, thus solving the problem of manually setting the number of layers required in traditional wavelet decomposition. N Question: Choose the db4 wavelet basis and set the initial number of decomposition levels. N =3, for standardized sequences P norm ( t Perform wavelet decomposition to obtain approximate components. A N (Low-frequency trends) and detail components D 1, D 2,..., D N (High-frequency fluctuations) satisfy Pnorm ( t )= A N + D 1+ D 2+...+ D N .

[0022] For adjacent detail components D k and D k+1 Calculate mutual information entropy E ( D k , D k+1 ): In the formula p ( x , y ) represents the joint probability density. p ( x ), p ( y ) represents the marginal probability density.

[0023] If it exists E ( D k , D k+1 () greater than the mutual information entropy threshold e To indicate the degree of scale aliasing, then N = N Repeat the decomposition and calculation by adding 1 until all adjacent components are reached. E All are less than or equal to the mutual information entropy threshold. e This indicates that the optimal number of layers has been obtained. N * .

[0024] Step 4: Based on the obtained number of layers, divide the high-frequency component set and low-frequency component set after wavelet decomposition.

[0025] High-frequency component set H Include D 1 to D K (in K less than or equal to N (Integers of * / 2) are used to capture short-term fluctuation characteristics, such as intraday changes in sunlight and the impact of gusts; low-frequency component sets L Include D K+1 to D N* as well as A N*Used to capture long-term trend characteristics, such as seasonal output characteristics and weekly cycles, ultimately yielding high-frequency components. H ( t )= D 1+ D 2+...+ D K Low-frequency components L ( t )= A N* + D K + D K+1 +...+ D N* .

[0026] Step 5: Use the weighted moving average method to predict high-frequency components.

[0027] Predicting high-frequency components using the weighted moving average method H pred ( t +1): In the formula w ti Indicates the first you The weights of historical data are determined by fitting the initial weight values ​​to the historical data, with more recent data having higher weights.

[0028] Step 6: Use linear regression to predict low-frequency components.

[0029] Predicting low-frequency components using linear regression L pred ( t +1): In the formula, i 0、 i 1 represents the coefficient of fit for linear regression. i 0、 i 1. Solve using the least squares method: .

[0030] Step 7: Combine the prediction results of high-frequency and low-frequency components to calculate the output value of the energy storage system at the next moment.

[0031] By integrating the prediction results of high-frequency and low-frequency components, the power output of the energy storage system in the power system at the next moment is calculated. P final ( t+ 1): when H pred (t +1)+ L pred ( t When +1)≥0: when H pred ( t +1)+ L pred ( t When +1) < 0: In the formula, P rate This is the rated power of the energy storage system.

[0032] It should be emphasized that the embodiments described in this invention are illustrative rather than limiting. Therefore, this invention includes, but is not limited to, the embodiments described in the specific implementation. Any other implementations derived by those skilled in the art based on the technical solutions of this invention are also within the scope of protection of this invention.

Claims

1. A control method for an energy storage system based on multi-scale feature fusion and dynamic adjustment, characterized in that: Includes the following steps: Step 1: Obtain historical renewable energy power generation and power load data, and obtain the net load sequence; Step 2: Data preprocessing and outlier correction, and zero-mean standardization of the net load series; Step 3: Optimize the number of layers in the net load sequence after zero-mean standardization based on wavelet decomposition using mutual information entropy; Step 4: Based on the obtained number of layers, divide the high-frequency component set and low-frequency component set after wavelet decomposition; Step 5: Predict high-frequency components using the weighted moving average method; Step 6: Use linear regression to predict low-frequency components; Step 7: Combine the prediction results of high-frequency and low-frequency components to calculate the output value of the energy storage system at the next moment.

2. The energy storage system control method based on multi-scale feature fusion and dynamic adjustment according to claim 1, characterized in that: The specific implementation method of step 1 is as follows: obtain the historical new energy power generation sequence. N ( t )( t =1,2,3..., T all , T all (Total length of historical data), power load data L ( t Then the net load sequence P ( t )for: 。 3. The energy storage system control method based on multi-scale feature fusion and dynamic adjustment according to claim 1, characterized in that: The specific implementation method of step 2 is as follows: The net load sequence is standardized to zero mean to eliminate the influence of dimensions. The net load sequence after zero mean standardization is... P norm ( t ): ； in, u P The arithmetic mean of the net load series. σ P is the standard deviation of the net load series.

4. The energy storage system control method based on multi-scale feature fusion and dynamic adjustment according to claim 1, characterized in that: The specific implementation method of step 3 is as follows: select the db4 wavelet basis, and set the initial decomposition level. N =3, for standardized sequences P norm ( t Perform wavelet decomposition to obtain approximate components. A N Low-frequency trends and detail components D 1, D 2,..., D N ,satisfy P norm ( t )= A N + D 1+ D 2+...+ D N ; For adjacent detail components D k and D k+1 Calculate mutual information entropy E ( D k , D k+1 ): ； in, p ( x , y ) represents the joint probability density. p ( x ), p ( y ) represents the marginal probability density; if it exists E ( D k , D k+1 () greater than the mutual information entropy threshold ε To indicate the degree of scale aliasing, then N = N Repeat the decomposition and calculation by adding 1 until all adjacent components are reached. E All are less than or equal to the mutual information entropy threshold. ε This indicates that the optimal number of layers has been obtained. N * .

5. The energy storage system control method based on multi-scale feature fusion and dynamic adjustment according to claim 1, characterized in that: The specific implementation method of step 4 is as follows: based on the obtained number of layers, divide the high-frequency component set and the low-frequency component set after wavelet decomposition, wherein the high-frequency component set... H Include D 1 to D K ,in K less than or equal to N Integers of * / 2 are used to capture short-term fluctuation characteristics; low-frequency component set L Include D K+1 to D N* as well as A N* Used to capture long-term trend characteristics, ultimately yielding high-frequency components. H ( t )= D 1+ D 2+...+ D K Low-frequency components L ( t )= A N* + D K + D K+1 +...+ D N* .

6. The energy storage system control method based on multi-scale feature fusion and dynamic adjustment according to claim 1, characterized in that: Step 5 uses the weighted moving average method to predict high-frequency components. H pred ( t The specific implementation method for +1 is as follows: ； in, w ti Indicates the first ti The weights of historical data are determined by fitting the initial weight values ​​to the historical data, with more recent data having higher weights.

7. The energy storage system control method based on multi-scale feature fusion and dynamic adjustment according to claim 1, characterized in that: Step 6 uses linear regression to predict low-frequency components. L pred ( t The specific implementation method for +1 is as follows: ； in, θ 0、 θ 1 represents the coefficient of fit for linear regression.

8. The energy storage system control method based on multi-scale feature fusion and dynamic adjustment according to claim 1, characterized in that: Step 7 integrates the prediction results of high-frequency and low-frequency components to calculate the output value of the energy storage system in the power system at the next moment. P final ( t+ The specific implementation method of 1) is as follows: when H pred ( t +1)+ L pred ( t When +1)≥0: ； when H pred ( t +1)+ L pred ( t When +1) < 0: ； in, P rate This is the rated power of the energy storage system.

Images