Load adaptive identification method based on multi-level feature library and reinforcement learning fusion

Abstract

The invention discloses a load adaptive identification method based on multi-level feature library and reinforcement learning fusion, and the method specifically comprises the steps: collecting load data and adjustment potential data of a user, and carrying out the preprocessing of the collected data, and obtaining multi-dimensional historical data; performing feature extraction on the multi-dimensional historical data; constructing a multi-level feature library based on the extracted features; preliminarily constructing a load pattern recognition model by training the multi-stage feature library; reinforcement learning is introduced, and self-adaptive learning and updating are carried out on the preliminarily constructed load mode recognition model through real-time data; and finally completing the identification of the load signal. According to the method, a data mining technology is utilized, a large-user multi-level feature library is constructed, the load mode is accurately recognized by training a preliminary load mode recognition model, introducing a reinforcement learning algorithm and combining real-time data for updating, and the recognition accuracy and real-time performance are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power systems, and in particular relates to a load adaptive identification method based on the fusion of a multi-level feature library and reinforcement learning. Background Art

[0002] With the continuous development of the power system and the gradual opening of the power market, load management plays an increasingly important role in the operation of the power system and power market transactions. As one of the core technologies of load management, load pattern recognition can achieve optimized dispatching and intelligent management of the power system by analyzing and identifying the power load pattern of users, thereby improving the operating efficiency and reliability of the power system.

[0003] Traditional load pattern recognition methods mainly rely on statistical analysis and feature extraction of historical data, usually using rule-based models or statistical learning methods. These methods can identify load patterns to a certain extent, but as time goes by and the external environment changes, the user's load pattern may change significantly. Traditional methods are difficult to reflect these changes in a timely manner due to the lack of real-time update mechanism. In order to solve the above problems, a load adaptive recognition method based on the fusion of multi-level feature library and reinforcement learning is proposed. Summary of the invention

[0004] The purpose of the present invention is to provide a load adaptive identification method based on the fusion of multi-level feature library and reinforcement learning. By using data mining technology, a multi-level feature library for large users is constructed. Through the training of preliminary load pattern recognition model, a reinforcement learning algorithm is introduced, and it is updated in combination with real-time data to accurately identify the load pattern and improve the recognition accuracy and real-time performance.

[0005] The technical solution adopted by the present invention is a load adaptive identification method based on the fusion of multi-level feature library and reinforcement learning, which is specifically implemented according to the following steps:

[0006] Step 1: Collect the user's load data and regulation potential data, and pre-process the collected data to obtain multi-dimensional historical data;

[0007] Step 2, extracting features from the multidimensional historical data in step 1;

[0008] Step 3, construct a multi-level feature library based on the features extracted in step 2;

[0009] Step 4: Preliminarily construct a load pattern recognition model by training the multi-level feature library;

[0010] Step 5: Introduce reinforcement learning to adaptively learn and update the load pattern recognition model initially constructed in step 4 through real-time data;

[0011] Step 6: Based on the load pattern recognition model that has completed adaptive learning and updating in Step 5, identify the load signal.

[0012] The features of the present invention also lie in that

[0013] The specific data preprocessing in Step 1 includes the following steps:

[0014] S11: Process the missing values and outliers in the collected data;

[0015] S12: Standardize the data that has been processed for missing values and outliers.

[0016] The missing value processing is specifically to scan each group of data to obtain the proportion of missing values in the data. If the proportion of missing values in a certain group of data exceeds 50%, then delete this group of data. If it does not exceed 50%, then determine whether there are three consecutive points of missing values in this group of data. If there are no more than three consecutive points of missing values, then perform linear interpolation to fill them, as shown in the following formula:

[0017]

[0018] In the formula, L i (t1) and L i (t2) are the known data at time points t1 and t2, and t1 < t < t2 is the time point of the missing value;

[0019] For more than three consecutive points of missing values, perform linear regression prediction, as shown in the following formula:

[0020] L i (t) = β0 + β1X1 + β2X2 + … + β p X p ;

[0021] In the formula, X1, X2, …, X p are the data of other features, and β0, β1, …, β p are the regression coefficients;

[0022] The outlier processing is to determine whether the deviation of the mean value of the data after missing value processing exceeds three times the standard. If it exceeds, then perform outlier processing, as shown in the following formula:

[0023]

[0024] In the formula, μ represents the data expected value, σ is the data standard deviation, and L i (t) represents the data at the i-th time point t;

[0025] If it does not exceed, then perform standardization processing, as shown in the following formula:

[0026]

[0027] In the formula, Represented as normalized data.

[0028] In step 2, the features are extracted based on the convolutional neural network and the LSTM network to extract the time-dependent features of the time load features and the regulation potential data respectively;

[0029] The convolutional neural network includes an input layer, three convolutional layers, a maximum pooling layer, a batch normalization layer, a flattening layer and a fully connected layer.

[0030] When the convolutional neural network extracts the time load feature, it specifically includes the following steps:

[0031] S21, the standardized data enters the input layer for processing;

[0032] S22, the data processed by the input layer is convolved by the convolution layer to generate a feature map;

[0033] S23, the feature map enters the maximum pooling layer to extract the maximum value and retain the most significant features;

[0034] S24, the output of the maximum pooling layer enters the batch normalization layer for standardization so that its mean is 0 and its variance is 1;

[0035] S25, the feature data after normalization is flattened into a one-dimensional vector by a flattening layer, and finally a fully connected layer performs linear transformation and nonlinear activation on the flattened feature vector to generate the final feature representation;

[0036] When the LSTM network extracts the time-dependent features of the regulatory potential data, the following steps are specifically included:

[0037] S26, the input layer receives the standardized regulatory potential data and performs time-dependent feature extraction;

[0038] S27, the gated attention mechanism is introduced in the forget gate. By using the sigmoid activation function and gated attention, a value between 0 and 1 is output, indicating the proportion of information that needs to be forgotten;

[0039] The input gate is based on X t and h t-1 The cell state is updated by combining the outputs of the forget gate and the input gate.

[0040] S28, using the dynamic weight adjustment mechanism, combined with the current cell state, modulated by the tanh activation function to generate the current hidden state;

[0041] S29, the fully connected layer further processes the output of the LSTM unit and generates the final feature representation through linear transformation and activation function.

[0042] Constructing a multi-level feature library in step 3 includes the following steps:

[0043] S31, based on the electricity consumption characteristics of typical industrial users, the primary database is divided into the machinery industry, steel industry, chemical industry, and nonferrous metal industry;

[0044] S32, based on the classification of the primary database, a secondary database is constructed according to the classification of the regulation potential capacity;

[0045] S33, classifying the load data by time period according to the performance of different load regulation potentials in different time periods, and used to construct a three-level database;

[0046] S34, classify the load data according to the response speed to construct a four-level database.

[0047] In step 4, training the multi-level feature library specifically includes the following steps:

[0048] S41, train the Transformer network based on the data in the multi-level database;

[0049] S42, training the GCN network based on the data in the multi-level database;

[0050] S43, the output of the Transformer network is used as the input of the graph neural network to extract the graph structure features, as shown in the following formula:

[0051] H (0) =H T ;

[0052] In the formula, H T is the output of the Transformer network.

[0053] Step 5 specifically includes the following steps:

[0054] S51, input the real-time data and map the real-time data into a high-dimensional space, as shown in the following formula:

[0055]

[0056] E 实时 =Embedding(X 实时 );

[0057] S52, based on the multi-head self-attention mechanism, captures the relationship between the various parts of the data by calculating the self-attention weights of real-time data, as shown in the following formula:

[0058]

[0059] In the formula, Q 实时 ,K 实时 ,V 实时 Respectively represent the query, key and value of real-time data, d k Represents the dimension of the key;

[0060] S53 performs a nonlinear transformation on the output of the self-attention mechanism and generates the final feature representation of the Transformer network, as shown in the following formula:

[0061] FFN 实时 (x) = max(0,xW1+b1)W2+b2;

[0062]

[0063] S54, the output of the Transformer network is used as the input of GCN, the relationship between nodes is captured through graph convolution operations, and then the node features are gradually updated through multiple layers of graph convolution layers. Finally, the output of the graph convolution layer is flattened into a one-dimensional vector and input into the fully connected layer for classification, as shown in the following formula:

[0064]

[0065] S55, using new real-time data and labels to update model parameters, as shown in the following formula:

[0066]

[0067] S56, based on the forecast results, identify the industry category, adjustment potential capacity, time period and response speed to which the data belongs.

[0068] The beneficial effects of the present invention are:

[0069] The load adaptive identification method based on the fusion of multi-level feature library and reinforcement learning of the present invention adopts feature extraction method combined with classification model to build multi-level feature library, which can extract high-level features from time series data, extract features of different levels based on different models, and form a similar integrated voting mechanism through the fusion of multi-level feature library and reinforcement learning to improve the accuracy and robustness of classification. Compared with the traditional problems of complex feature extraction, poor generalization ability of single classification model and low classification accuracy, the present invention reduces the complexity of feature engineering and ensures the accuracy and stability of load pattern recognition.

[0070] Traditional load pattern recognition methods have poor adaptability and cannot reflect the dynamic changes of load patterns in a timely manner. The present invention introduces reinforcement learning, which enables the model to perform adaptive learning and dynamic optimization based on real-time data, improves the model's real-time response capability to changes in load patterns, and enhances the accuracy and real-time performance of recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a flow chart of a load adaptive identification method based on the fusion of multi-level feature library and reinforcement learning of the present invention;

[0072] Figure 2 It is a flow chart of data processing and feature extraction in the load adaptive identification method based on the fusion of multi-level feature library and reinforcement learning of the present invention;

[0073] Figure 3 It is a schematic diagram of the structure of a multi-level feature library in the load adaptive identification method based on the fusion of a multi-level feature library and reinforcement learning in the present invention;

[0074] Figure 4 It is a structural schematic diagram of a load pattern recognition model in a load adaptive recognition method based on the fusion of a multi-level feature library and reinforcement learning in the present invention. DETAILED DESCRIPTION

[0075] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0076] The present invention proposes a load adaptive identification method based on the fusion of multi-level feature library and reinforcement learning. First, the multi-dimensional raw data from the acquisition system and the marketing system are preprocessed, including processing missing values ​​and outliers, and the processed data are standardized. Then, based on different feature extraction methods, the multi-dimensional features are analyzed. Considering the feature differences of different types of data, convolutional neural networks and improved LSTM networks are used for feature extraction to construct a multi-level feature library. Next, a Transformer network is designed in combination with a graph neural network, and the data in the established feature library is used for training to preliminarily construct a load pattern recognition model. Finally, through reinforcement learning, the intelligent agent interacts with the environment, and real-time data is used as environmental feedback to continuously learn and optimize decision-making strategies, and finally output recognition results.

[0077] Example 1

[0078] like Figure 1 As shown, the load adaptive identification method based on the fusion of multi-level feature library and reinforcement learning of the present invention is implemented in the following steps:

[0079] Step 1: Collect the user's load data and regulation potential data, and pre-process the collected data to obtain multi-dimensional historical data; the data pre-processing specifically includes the following steps:

[0080] S11, processing missing values ​​and outliers on the collected data;

[0081] S12, standardize the data after missing values ​​and outliers are processed.

[0082] Step 2, extracting features from the multidimensional historical data in step 1; specifically, extracting features based on a convolutional neural network and an LSTM network to extract time-dependent features of time load features and regulation potential data respectively;

[0083] The convolutional neural network includes an input layer, three convolutional layers, a maximum pooling layer, a batch normalization layer, a flattening layer and a fully connected layer.

[0084] The present invention adopts a three-layer convolutional network to extract load features: the convolution kernel size of the first convolution layer is [3, 1], the number of channels is 32, and it is connected to the batch normalization layer through the first activation layer; the convolution kernel size of the second convolution layer is [3, 1], the number of channels is 64, and it is connected to the batch normalization layer through the second activation layer; the convolution kernel size of the third convolution layer is [3, 1], the number of channels is 128, and it is connected to the maximum pooling layer through the third activation layer; after the maximum pooling layer, the output is standardized through the batch normalization layer. Then, the multidimensional feature map is flattened into a one-dimensional vector through the flattening layer, and finally the flattened feature vector is linearly transformed and nonlinearly activated through the fully connected layer to generate the final feature representation.

[0085] The LSTM network is used to extract the time-dependent features of the regulatory potential data: First, the input layer receives the standardized regulatory potential data and prepares for time-dependent feature extraction. The LSTM network dynamically adjusts the weights in the forget gate, input gate, and output gate by introducing a gated attention mechanism, a hybrid activation function, and a dynamic weight adjustment mechanism, respectively, to accurately control the forgetting, updating, and output of information.

[0086] Step 3, construct a multi-level feature library based on the features extracted in step 2;

[0087] Step 4: Preliminarily construct a load pattern recognition model by training the multi-level feature library;

[0088] Step 5: Introduce reinforcement learning to adaptively learn and update the load pattern recognition model initially constructed in step 4 through real-time data;

[0089] Step 6: Based on the load pattern recognition model that has completed adaptive learning and updating in step 5, the load signal is recognized.

[0090] Example 2

[0091] The load adaptive recognition method based on the fusion of a multi-level feature library and reinforcement learning is specifically implemented according to the following steps:

[0092] Step 1, as Figure 2 shown, collect the load data and regulation potential data of the user, and preprocess the collected data to obtain multi-dimensional historical data; the regulation potential data includes adjustable potential data, adjustable response capacity, adjustable response time period, and adjustable response speed.

[0093] The data preprocessing specifically includes missing value, outlier, and normalization processing. During the process of processing missing values, the proportion of missing values in the data needs to be considered. First, scan each group of data to obtain the proportion of missing values in the data. If the proportion of missing values in a certain group of data exceeds 50%, it is considered that this group of data is not sufficient to support the analysis, so this feature is deleted. If it does not exceed 50%, appropriate methods need to be taken to fill it. Determine whether there are three consecutive points of missing values in this group of data. If there are no more than three consecutive points of missing values, linear interpolation filling is performed, as shown in the following formula:

[0094]

[0095] In the formula, L i (t1) and L i (t2) are the known data at time points t1 and t2, and t1 < t < t2 is the time point of the missing value;

[0096] For data with periodic changes, data with more than three consecutive points of missing values and a missing value proportion not exceeding 50% are predicted using a linear regression model, as shown in the following formula:

[0097] L i (t) = β0 + β1X1 + β2X2 + … + β p X p ;

[0098] In the formula, X1, X2, …, X p are the data of other features, and β0, β1, …, β p are the regression coefficients;

[0099] To improve the reliability of the data and the accuracy of the analysis, the outlier processing uses the three-standard-deviation method. According to the statistical principle of the normal distribution, under the normal distribution, 99.7% of the data points will fall within three standard deviations of the mean. Therefore, data points deviating more than three standard deviations from the mean can be reasonably regarded as outliers, as shown below:

[0100]

[0101] In the formula, μ represents the data expectation value, σ is the data standard deviation, Li (t) represents the data at the i-th time point t;

[0102] By calculating the difference between each data point and the mean and comparing it with three times the standard deviation, outliers can be effectively identified and treated as missing values. If it does not exceed, it is standardized to eliminate the dimensional differences between features, as shown below:

[0103]

[0104] In the formula, Represented as normalized data.

[0105] Step 2: Extract features from the multidimensional historical data in step 1.

[0106] The time load feature in step 2 is extracted based on the convolutional neural network (CNN). A convolutional neural network is a deep learning model that consists of multiple convolutional layers and pooling layers stacked alternately, and outputs through the fully connected layer at the end to predict the input parameters.

[0107] The specific steps include:

[0108] S21, the standardized data enters the input layer for processing, as shown in the following formula:

[0109]

[0110] Where X is the input data;

[0111] S22, the convolution layer is the core of the network. All convolution kernels extract the features of the input samples through weight sharing. Multiple convolution kernels are applied to perform convolution operations on the input data. The convolution layer convolves the input data through a sliding window. The data processed by the input layer is convolved by three convolution layers to generate a feature map, as shown in the following formula:

[0112]

[0113] In the formula, h i,j is the output feature map of the i-th row and j-th column of the convolutional layer, X is the data matrix of the input layer, and W m is the weight of the convolution kernel at the mth position, b i is the bias term of the i-th convolution kernel.

[0114] The activation layer introduces nonlinearity through the ReLU (Rectified Linear Unit) activation function, which effectively solves the gradient vanishing problem and improves training speed and model performance.

[0115] a i,j =ReLU(hi,j )=max(0,h i,j );

[0116] In the formula, a i,j is the output of the activation layer.

[0117] S23, the feature map enters the maximum pooling layer to extract the maximum value and retain the most significant features. The pooling layer reduces the dimension of the feature vector to prevent overfitting during prediction. The present invention adopts the maximum pooling method to retain the most significant features by taking the maximum value in the pooling window. In order to prevent the loss of important feature information, only one maximum pooling layer is designed, as shown in the following formula:

[0118] p i,j =max{a i+m,j+n ∣0≤m,n <s};

[0119] In the formula, p i,j is the output feature map of the pooling layer, and s is the size of the pooling window

[0120] S24, the output of the maximum pooling layer enters the batch normalization layer for standardization, making its mean 0 and variance 1, thereby reducing internal covariate transfer and accelerating model convergence, as shown in the following formula:

[0121]

[0122] In the formula, For normalized data, ∈ is a very small constant;

[0123] S25, the normalized feature data is flattened into a one-dimensional vector by the flattening layer so as to be input into the fully connected layer for further processing, as shown in the following formula:

[0124] f = flatten(p);

[0125] Where f is the expanded eigenvector;

[0126] Finally, the fully connected layer performs linear transformation and nonlinear activation on the flattened feature vector to generate the final feature representation, as shown in the following formula:

[0127] o=σ(Wf+b);

[0128] Where o is the output vector of the fully connected layer.

[0129] Furthermore, the present invention adopts the improved Long short-term memory network (LSTM) for potential data extractionThe network is good at processing and predicting long-term dependency problems in time series data. By introducing a gating mechanism, it can effectively capture long-term and short-term dependencies in time series. In order to solve the limitations of traditional LSTM, the present invention enhances its feature extraction capability through an improved gating setting, which specifically includes the following steps:

[0130] S26, the input layer receives the standardized regulatory potential data and performs time-dependent feature extraction, as shown in the following formula:

[0131]

[0132] Where X is the data input matrix, is the standardized potential adjustment data at the i-th time point t, and N is the total number of data points.

[0133] At each time step of training, the LSTM network first accepts the input at the current time and the hidden state at the previous time.

[0134] S27, the gated attention mechanism is introduced into the forget gate to determine which parts of the cell state need to be discarded. By using the sigmoid activation function and gated attention, a value between 0 and 1 is output, indicating the proportion of information that needs to be forgotten, as shown in the following formula:

[0135] f t =σ(W f ·[h t-1 ,X t ]+b f +Attention(h t-1 ,X t ));

[0136] In the formula, f t is the output of the forget gate, W f is the weight matrix, h t-1 is the hidden state of the previous time step, X t is the current input, b f Bias term, the attention mechanism is used to dynamically adjust the weights.

[0137] The input gate is based on X t and h t-1 The mixed activation function (sigmoid+ReLU) is used to output a value between 0 and 1, indicating the proportion of information that needs to be updated, as shown in the following formula:

[0138] i t =σ(W i ·[h t-1 ,X t ]+b i )·ReLU(Wi ′·[h t-1 ,X t ]+b i ′);

[0139] In the formula, i t is the output of the input gate, W i and W i ′ is the weight matrix, b i and b i ′ is the bias term. Candidate memory is the new candidate information generated by the tanh activation function, to be selectively added to the cell state C t As shown in the following formula:

[0140]

[0141] In the formula, is a candidate memory, W C is the weight matrix.

[0142] Combine the outputs of the forget gate and the input gate to update the cell state. The forget gate controls the proportion of old information forgotten, and the input gate controls the proportion of new information added, as shown in the following formula.

[0143]

[0144] In the formula, C t is the current cell state, and * represents the element-wise product. The output gate determines how much information in the current cell state will be used as the current output or hidden state.

[0145] S28, using the dynamic weight adjustment mechanism, combined with the current cell state, modulated by the tanh activation function, generates the current hidden state, as shown in the following formula:

[0146] o t =σ(W o ·[h t-1 ,X t ]+b o )+Attention(C t ,h t-1 );

[0147] h t =o t *tanh(C t );

[0148] In the formula, o t is the output of the output gate, and the attention mechanism is used to dynamically adjust the output weight.

[0149] S29, the fully connected layer further processes the output of the LSTM unit and generates the final feature representation through linear transformation and activation function, as shown in the following formula:

[0150] o=σ(Wh T + b);

[0151] Where o is the output vector of the fully connected layer.

[0152] Step 3: construct a multi-level feature library based on the features extracted in step 2; constructing the multi-level feature library includes the following steps:

[0153] S31, such as Figure 3 As shown, based on the electricity consumption characteristics of typical industrial users, the primary database is divided into the machinery industry, steel industry, chemical industry, and nonferrous metal industry;

[0154] The industry category of industrial load is the first level classification of the feature library. According to different industrial categories, the load data is divided into machinery, chemicals, steel and non-ferrous metals. This classification helps to capture the specific load characteristics of each industry and improve the efficiency and accuracy of feature extraction.

[0155] For the machinery industry, the electricity demand at night is quite different from that during the day. The electricity demand from 0:00 to 7:00 is relatively low, and then the load level increases rapidly over time until it reaches the highest load at 1:00 p.m., then slowly decreases at 20:00, and then basically stabilizes. For the chemical industry, its electricity consumption has obvious peaks and valleys: the peaks are mainly at 5:00-8:00, 11:00-16:00 and 19:00-21:00, and the troughs are mainly at 3:00-5:00, 8:00-11:00, 17:00-18:00 and 22:00-24:00. The electricity demand of the steel industry is lower at night, and the load demand fluctuates less during the period from 0:00 to 6:00, which is generally low, with the lowest at 6:00, and then gradually increases to 14:00 to reach the highest, and remains until 20:00 before starting to decline. The load demand of the nonferrous metals industry does not change much during the day and basically remains at the same level in different time periods. The main reason is that it requires uninterrupted production, so most equipment requires 24-hour power supply.

[0156] S32, based on the classification of the primary database, a secondary database is constructed according to the classification of the regulation potential capacity;

[0157] The capacity of regulation potential is the second level classification of the feature library. According to different load regulation potentials, the load data is further subdivided. The load regulation potential is divided into three categories: large, medium, and small, representing large capacity, medium capacity, and small capacity that can be adjusted.

[0158] S33, classifying the load data by time period according to the performance of different load regulation potentials in different time periods, and used to construct a three-level database;

[0159] Different load regulation potentials show different performances in different time periods. By classifying the load data by time period, it is specifically divided into: 0:00-6:00: The electricity demand is relatively low, the load fluctuation is small, and it is suitable for nighttime production and maintenance operations. 6:00-12:00: The electricity demand begins to increase, especially when industrial and commercial activities become active, and the load gradually rises. 12:00-18:00: The electricity demand is at its peak, and the load reaches the highest point of the day. 18:00-24:00: The electricity demand gradually decreases, but it still remains at a high level, especially during the peak hours of household and commercial activities.

[0160] S34, classifying the load data according to the response speed to construct a four-level database;

[0161] The speed of response of different loads varies significantly, and the fast-response load has important strategic significance in load management. Specifically, they are divided into: fast response (0-5 minutes), medium response (5-15 minutes), slow response (15-30 minutes), and ultra-slow response (more than 30 minutes).

[0162] Step 4: Preliminarily construct a load pattern recognition model by training the multi-level feature library;

[0163] Step 5: Introduce reinforcement learning to adaptively learn and update the load pattern recognition model initially constructed in step 4 through real-time data;

[0164] Step 6: Based on the load pattern recognition model that has completed adaptive learning and updating in step 5, the load signal is recognized.

[0165] Example 3

[0166] On the basis of Example 2, in step 4 of the load adaptive identification method based on the fusion of multi-level feature library and reinforcement learning of the present invention, training the multi-level feature library specifically includes the following steps:

[0167] S41, such as Figure 4 As shown, the Transformer network is trained based on the data in the multi-level database;

[0168] The Transformer model is mainly composed of position encoding, encoder-decoder, residual connection and layer normalization modules. In the Transformer model, position encoding plays a vital role. Since the Transformer model mainly relies on the self-attention mechanism to process the input sequence, it does not have the ability to process the position information of elements in the sequence. Therefore, position encoding is introduced to embed the position information in the input sequence. The position encoding is usually a matrix with the same dimension as the input sequence, and the position information is encoded by using sine and cosine functions, as shown in the following formula:

[0169]

[0170] Where: PE represents the position encoding. pos represents the position of the current character in the input sequence. i is the index of the dimension, indicating whether the dimension currently being processed is an even position or an odd position. modsl Represents the dimension of the model, that is, the dimension size of the position encoding vector. The historical input data in the database is accepted as a time series, as shown in the following formula:

[0171]

[0172] In the formula, X represents the input data, and Represent the standardized load data and regulation potential data respectively. In order to improve the expressiveness of the model, the embedding layer maps the low-dimensional input data to a high-dimensional space. The input data is mapped to a high-dimensional space as shown in the following formula:

[0173] E = Embedding(X);

[0174] Where E represents the embedded data. Each encoder-decoder consists of two sublayers: a self-attention layer and a feedforward neural network. The multi-head self-attention mechanism can calculate attention in different subspaces in parallel and capture richer features. Through parallel processing of multiple self-attention heads, the relationship between the various parts of the data is captured, as shown in the following formula:

[0175]

[0176] Where Q, K, and V represent query, key, and value respectively. k Represents the dimension of the key. When multiple self-attention layers are stacked together, each layer receives the output of the previous layer as input and produces a new output, which is integrated into the output through nonlinear transformation and attention mechanism, as shown in the following formula:

[0177] MultiHeadQ,K,V=Concathead1,…,head n W O ;

[0178]

[0179] In the formula, each head i It is the head of the self-attention mechanism. The output of the self-attention mechanism is nonlinearly transformed through the feedforward neural network. The output layer further processes the result of the feedforward neural network to generate the final feature representation, as shown in the following formula:

[0180] FFN(x)=ReLU(0,xW1+b1)W2+b2;

[0181] H T =OutputLayer(FFN);

[0182] Where W1, W2, b1, b2 are the weights and bias terms of the network, ReLU is the activation function, and H T Represents the output of the Transformer network.

[0183] S42, based on the data in the multi-level database, train the GCN network as shown in the following formula:

[0184]

[0185] In the formula, represents the adjacency matrix with self-loops added, express The degree matrix, H (l) represents the node features of the lth layer, W (l) represents the weight of the lth layer, and σ represents the activation function. The graph convolution layer aggregates the information of neighboring nodes and updates the node features, as shown in the following formula:

[0186]

[0187] O=σ(WH final + b);

[0188] In the formula, H final The node features of the last layer, the fully connected layer maps the high-dimensional features to the target output space.

[0189] S43, the output of the Transformer network is used as the input of the graph neural network to extract the graph structure features, as shown in the following formula:

[0190] H (0) =H T ;

[0191] In the formula, H T is the output of the Transformer network.

[0192] In this way, the long-distance dependency features extracted by the Transformer network can be further processed by the graph neural network to capture the relationship between nodes. The model is trained using the data in the established feature library. The loss function used is the cross-entropy loss.

[0193]

[0194] In the formula, y i represents the true label, Represents the predicted probability.

[0195] In step 5, reinforcement learning is introduced, and the load pattern recognition model constructed is adaptively learned and updated through real-time data. Specifically, the following steps are included:

[0196] S51, input the real-time data and map the real-time data into a high-dimensional space for subsequent processing, as shown in the following formula:

[0197]

[0198] E 实时 =Embedding(X 实时 );

[0199] S52, based on the multi-head self-attention mechanism, captures the relationship between the various parts of the data by calculating the self-attention weights of real-time data, as shown in the following formula:

[0200]

[0201] In the formula, Q 实时 ,K 实时 ,V 实时 Respectively represent the query, key and value of real-time data, d k Represents the dimension of the key;

[0202] S53 performs a nonlinear transformation on the output of the self-attention mechanism and generates the final feature representation of the Transformer network, as shown in the following formula:

[0203] FFN 实时 (x) = max(0,xW1+b1)W2+b2;

[0204]

[0205] S54, the output of the Transformer network is used as the input of GCN, the relationship between nodes is captured through graph convolution operations, and then the node features are gradually updated through multiple layers of graph convolution layers. Finally, the output of the graph convolution layer is flattened into a one-dimensional vector and input into the fully connected layer for classification, as shown in the following formula:

[0206]

[0207] S55, using new real-time data and labels to update model parameters, as shown in the following formula:

[0208]

[0209] S56, based on the forecast results, identify the industry category, adjustment potential capacity, time period and response speed to which the data belongs.

Claims

1. A load adaptive identification method based on the fusion of multi-level feature library and reinforcement learning, characterized in that: Follow the steps below to implement it: Step 1: Collect the user's load data and regulation potential data, and pre-process the collected data to obtain multi-dimensional historical data; Step 2, extracting features from the multidimensional historical data in step 1; Step 3, construct a multi-level feature library based on the features extracted in step 2; Step 4: Preliminarily construct a load pattern recognition model by training the multi-level feature library; Step 5: Introduce reinforcement learning to adaptively learn and update the load pattern recognition model initially constructed in step 4 through real-time data; Step 6: Based on the load pattern recognition model that has completed adaptive learning and updating in step 5, the load signal is recognized.

2. The load adaptive identification method based on the fusion of multi-level feature library and reinforcement learning according to claim 1 is characterized in that: The data preprocessing in step 1 specifically includes the following steps: S11, processing missing values ​​and outliers on the collected data; S12, standardize the data after missing values ​​and outliers are processed.

3. The load adaptive identification method based on the fusion of multi-level feature library and reinforcement learning according to claim 2 is characterized in that: The missing value processing is specifically to scan each group of data to obtain the missing value ratio of the data. If the missing value ratio of a group of data exceeds 50%, the group of data is deleted. If it does not exceed 50%, it is determined whether there are three consecutive points of missing values ​​in the group of data. If no more than three consecutive points are missing values, linear interpolation is performed to fill the missing values, as shown in the following formula: where L i (t1) and L i (t2) are known data at time points t1 and t2, and t1 < t < t2 is the time point of the missing value; For more than three consecutive points with missing values, linear regression prediction is performed as shown below: L i (t)=β0+β1X1+β2X2+…+β p X p ; In the formula, X1, X2, …, X p is the data of other features, β0,β1,…,β p is the regression coefficient; The outlier processing is to determine whether the mean deviation of the data after missing value processing exceeds three times the standard. If it exceeds, outlier processing is performed, as shown in the following formula: In the formula, μ represents the expected value of the data, σ is the standard deviation of the data, and L i (t) represents the data at the i-th time point t; If it does not exceed, standardization is performed as shown below: In the formula, Represented as normalized data.

4. The load adaptive identification method based on the fusion of multi-level feature library and reinforcement learning according to claim 1 is characterized in that: In the step 2, the features are extracted based on the convolutional neural network and the LSTM network to extract the time load features and the time dependency features of the regulation potential data respectively; The convolutional neural network includes an input layer, three convolutional layers, a maximum pooling layer, a batch normalization layer, a flattening layer and a fully connected layer.

5. The load adaptive identification method based on the fusion of multi-level feature library and reinforcement learning according to claim 4 is characterized in that: When the convolutional neural network extracts the time load feature, the following steps are specifically included: S21, the standardized data enters the input layer for processing; S22, the data processed by the input layer is convolved by the convolution layer to generate a feature map; S23, the feature map enters the maximum pooling layer to extract the maximum value and retain the most significant features; S24, the output of the maximum pooling layer enters the batch normalization layer for standardization so that its mean is 0 and its variance is 1; S25, the normalized feature data is flattened into a one-dimensional vector by a flattening layer, and finally the fully connected layer performs linear transformation and nonlinear activation on the flattened feature vector to generate the final feature representation.

6. The load adaptive identification method based on the fusion of multi-level feature library and reinforcement learning according to claim 4 is characterized in that: When the LSTM network extracts the time-dependent features of the regulatory potential data, the following steps are specifically included: S26, the input layer receives the standardized regulatory potential data and performs time-dependent feature extraction; S27, the gated attention mechanism is introduced in the forget gate. By using the sigmoid activation function and gated attention, a value between 0 and 1 is output, indicating the proportion of information that needs to be forgotten; The input gate is based on X t and h t-1 The cell state is updated by combining the outputs of the forget gate and the input gate. S28, using the dynamic weight adjustment mechanism, combined with the current cell state, modulated by the tanh activation function to generate the current hidden state; S29, the fully connected layer further processes the output of the LSTM unit and generates the final feature representation through linear transformation and activation function.

7. The load adaptive identification method based on the fusion of multi-level feature library and reinforcement learning according to claim 1 is characterized in that: The step 3 of constructing a multi-level feature library includes the following steps: S31, based on the electricity consumption characteristics of typical industrial users, the primary database is divided into the machinery industry, steel industry, chemical industry, and nonferrous metal industry; S32, based on the classification of the primary database, a secondary database is constructed according to the classification of the regulation potential capacity; S33, classifying the load data by time period according to the performance of different load regulation potentials in different time periods, and used to construct a three-level database; S34, classify the load data according to the response speed to construct a four-level database.

8. The load adaptive identification method based on the fusion of multi-level feature library and reinforcement learning according to claim 1 is characterized in that: The training of the multi-level feature library in step 4 specifically includes the following steps: S41, train the Transformer network based on the data in the multi-level database; S42, training the GCN network based on the data in the multi-level database; S43, the output of the Transformer network is used as the input of the graph neural network to extract the graph structure features, as shown in the following formula: H (0) =H T ; In the formula, H T is the output of the Transformer network.

9. The load adaptive identification method based on the fusion of multi-level feature library and reinforcement learning according to claim 1 is characterized in that: The step 5 specifically comprises the following steps: S51, input the real-time data and map the real-time data into a high-dimensional space, as shown in the following formula: E 实时 =Embedding(X 实时 ); S52, based on the multi-head self-attention mechanism, captures the relationship between the various parts of the data by calculating the self-attention weights of real-time data, as shown in the following formula: In the formula, Q 实时 ,K 实时 ,V 实时 Respectively represent the query, key and value of real-time data, d k Represents the dimension of the key; S53 performs a nonlinear transformation on the output of the self-attention mechanism and generates the final feature representation of the Transformer network, as shown in the following formula: FFN 实时 (x)=max(0,xW1+b1)W2+b2; S54, the output of the Transformer network is used as the input of GCN, the relationship between nodes is captured through graph convolution operations, and then the node features are gradually updated through multiple layers of graph convolution layers. Finally, the output of the graph convolution layer is flattened into a one-dimensional vector and input into the fully connected layer for classification, as shown in the following formula: S55, using new real-time data and labels to update model parameters, as shown in the following formula: S56, based on the forecast results, identify the industry category, adjustment potential capacity, time period and response speed to which the data belongs.

Images