Industrial load portrait and situation prediction system and method based on multi-time scale characteristics

Abstract

The invention provides an industrial load portrait and situation prediction system and method based on multi-time scale characteristics, and relates to the technical field of industrial load and situation prediction. According to the method, diversified industrial loads are acquired by acquiring power utilization information and equipment characteristics of users in an industrial park, an industrial load portrait is established accordingly, and the diversified industrial loads are analyzed by fusing multi-time scale sequence characteristics with industrial load portrait characteristics and combining a double-layer multi-head self-attention mechanism. The feature multi-head self-attention layer focuses important features and excavates input feature sequence relevance, and the time sequence multi-head self-attention layer excavates periodic trend features and local fluctuation features of industrial loads by capturing time sequence relevance and strengthening information expression of important feature variables and key time points; the method can achieve the multi-time-scale prediction of the energy consumption situation of the load of the industrial park, remarkably increases the flexible regulation and control space of the industrial park, and improves the digital perception and flexible regulation and control level of a demand side.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial load and situation prediction, and in particular to a system and method for industrial load profiling and situation prediction based on multi-time scale characteristics. Background Art

[0002] Industrial parks cover a large number of diversified loads. Deducing the regulation potential of diversified load clusters and carrying out precise regulation are key links in achieving friendly interaction between the new power system and the load. The excellence of power system planning and operation depends to a large extent on the accurate prediction of loads. Accurate and reliable load prediction helps avoid decisions that worsen the operation of the power grid, and is conducive to industrial users and power departments to grasp load fluctuations and power consumption in a timely manner, and plan based on the prediction results, maintain system stability and reliability, and optimize production processes and costs.

[0003] At present, industrial load forecasting research is mainly based on traditional statistical models and machine learning methods. However, industrial electricity demand is affected by many factors, and some factors in typical forecasting models are not always applicable. For example, seasonality, calendar events, and weather dependence are important in electricity demand curves, but they are inconsistent between industrial sites. Household and industrial loads follow different patterns, and typical models are difficult to adapt to all industrial sites. When conditions permit, establishing industrial load profiles can improve forecast accuracy and reliability.

[0004] In the study of industrial load profiling, Song Yi et al. constructed a framework for evaluating the response potential of flexible resources in "A Method for Evaluating the Response Potential of Flexible Resources for Regional Grids" Journal of Electric Power Systems and Automation, 2023, 35(08) , but it cannot intuitively describe the size of the response capacity available in each time period and cannot provide a direct reference for real-time demand response events. "Xu Chengsi proposed an evaluation system for the comprehensive demand response capacity of industrial users in "Research on Comprehensive Demand Response in Industrial Parks" Zhejiang University, 2021 , but did not consider the impact of the scheduling period on the response characteristics of electricity consumption behavior. Zheng Ruonan et al. constructed an incentive-based demand response model in "Incentive-based Demand Response Model and Evaluation Considering Uncertainty of Residential User Participation" Automation of Electric Power Systems, 2022, 46(08) to screen users with high regulation potential, but the modeling of industrial equipment was simplified and the differences in equipment were not fully considered. Zhang Haijing et al. evaluated the response capacity of industrial users in "Research on Economic Optimization Modeling and Regulation Potential of Power Generation and Consumption of Typical High-Energy-Consuming Industrial Users" Electricity Demand Side Management, 2021, 23(02) to cope with the uncertainty of residents' response behavior, but the method is mainly applicable to a single type of load and has limitations.

[0005] In the study of industrial load forecasting, HUANG SJ et al. used bispectral analysis to optimize the ARMA model and improve the forecasting performance in “Short-term load forecasting including non-Gaussian process considerations through ARMA model identification” [J]. IEEE Transactions on Power Systems, 2003, 18(2): 673-679. BRACALE A et al. compared the univariate and multivariate methods in “Comparison of univariate and multivariate methods for probabilistic forecasting of industrial loads” (Proceedings of the 5th International Symposium on Environmentally Friendly Energy and Applications) IEEE Press, 2018: 1-6. HONG T et al. proposed a framework based on quantile regression to build a forecasting model in “Probabilistic Power Load Forecasting: A Tutorial Review” [J]. International Journal of Forecasting, 2016, 32(3): 914-938. The above methods are simple, efficient and highly interpretable, but it is difficult to describe the nonlinearity and randomness of industrial loads. WANG YY et al. combined VMD and XGBoost algorithms for prediction in "Short-term load forecasting for industrial customers based on SVMD and XGBoost" [J]. International Journal of Electrical Power & Energy Systems, 2021, 129: 106830.", but the variable association evaluation depends on expert knowledge.

[0006] In summary, in terms of building industrial load portraits, although many studies have quantified the response capabilities, they are limited to a single type of load and lack a comprehensive analysis of diversified loads. In addition, research on situation prediction often fails to take into account the randomness of industrial load fluctuations and the regular characteristics of industrial users' electricity consumption behavior. In other words, existing research still needs further exploration and improvement in terms of model applicability, personalized demand analysis, user behavior complexity, and objectivity of data processing.

[0007] Based on this, the present invention is proposed. Summary of the invention

[0008] In order to solve at least one problem existing in the background technology, the purpose of the present invention is to provide an industrial load profiling and situation prediction system and method based on multi-time scale characteristics, in order to realize multi-time scale prediction of the energy consumption situation of industrial park load, which can explore the periodic trend characteristics and local fluctuation characteristics of industrial load, take into account the randomness of industrial load fluctuations and the regular characteristics of industrial users' electricity consumption behavior, and can be applied to different industrial load scenarios, significantly increase the flexible regulation space of industrial parks, and improve the digital perception and flexible regulation level of the demand side.

[0009] To achieve the above object, the present invention provides the following technical solutions: In a first aspect, the present invention provides an industrial load portrait and situation prediction method based on multi-time scale features, including the following: S1: Collecting electricity consumption information and equipment characteristics of industrial park users; S2: Based on the collected electricity consumption information and equipment characteristics of industrial park users, introducing a graph attention network to construct an industrial load portrait model to obtain industrial load portrait characteristics; S3: Based on the collected electricity consumption information and equipment characteristics of industrial park users, constructing a multi-time scale sequence feature input model to obtain multi-time scale sequence characteristics; S4: Constructing a feature extraction model based on a feature multi-head self-attention layer, integrating the industrial load portrait characteristics with the multi-time scale sequence characteristics through the feature extraction model, using attention to weight each feature, quantifying the correlation of each feature and focusing on important features; S5: Constructing a prediction model based on a time series multi-head self-attention layer, inputting the data output by the feature extraction model into the prediction model for training, and the trained prediction model is used to output the industrial load situation prediction result.

[0010] In the first aspect, the present invention provides a preferred scheme, S2 specifically includes: S21 constructing a load prior behavior association graph: according to the collected electricity consumption information and equipment characteristics of industrial park users, a set of load nodes of various load behaviors and load numbers of industrial park users and a set of directed edges of transfer relationships and association numbers between different load behaviors are obtained, and a load prior behavior association graph represented by a directed graph is constructed, and the prior transfer probability between any load behaviors in the load prior behavior association graph is calculated, and a load prior behavior transfer matrix is ​​constructed according to the prior transfer probability; S22 improved load behavior association graph extraction of graph attention network: a shared weight parameter matrix is ​​introduced to perform input feature transformation on the feature vectors of the load nodes in the load prior behavior transfer matrix as the input feature vector of the graph attention network, and the graph attention network outputs the attention coefficients of each adjacent node, and the feature vectors of the corresponding load nodes in the load prior behavior association graph in S21 are optimized by the attention coefficients, so as to realize the improved load behavior association graph extraction of the graph attention network.

[0011] In a first aspect, the present invention provides a preferred solution. In S21, binary smoothing is performed on the calculated priori transfer probabilities between any load behaviors, and the load priori behavior transfer matrix is ​​constructed based on the priori transfer probabilities after binary smoothing.

[0012] The present invention provides a preferred solution in the first aspect, S2 also includes: using the Softmax activation function to normalize the attention coefficient of each adjacent node, and optimizing the feature vector of the corresponding load node through the normalized attention coefficient.

[0013] More preferably, the optimization method is: performing weighted summation on the feature vectors of the corresponding load nodes through the normalized attention coefficients of each adjacent node to obtain a new output feature vector of the corresponding load node.

[0014] In a first aspect, the present invention provides a preferred solution, in S3, a multi-time scale sequence feature input model is constructed using wavelet transform or Fourier transform, which is used to decompose time series data into sequence data of different time scales and extract features from the decomposition results of different time scales.

[0015] The present invention provides a preferred solution in the first aspect, S4 specifically includes the following: S41: Obtaining a weighted quantized feature matrix: Weighting the input feature sequence matrix in the feature multi-head self-attention layer to obtain the weight of any feature sequence, thereby obtaining a weighted quantized feature matrix; S42: Capturing important features at multiple scales and angles: Obtaining n Query matrices from the input feature sequence matrix through the feature multi-head self-attention layer, and after multi-head self-attention weighted quantization, obtaining a weighted matrix, thereby realizing multi-scale and multi-angle capture of important features.

[0016] The present invention provides a preferred solution in the first aspect. In S5, the prediction model uses a BiGRU neural network to train the input data; the BiGRU network includes an input layer, a hidden layer and an output layer.

[0017] More preferably, at any moment, the hidden layer state output by the BiGRU neural network is jointly determined by the output results of the forward GRU and the backward GRU at that moment, and the weighting matrix obtained in S42 is input into the BiGRU neural network for feature training to obtain an updated hidden layer state; the temporal multi-head self-attention layer is applied to the output end of the BiGRU network, and the temporal multi-head self-attention layer quantizes the weights of the hidden layer state output by the BiGRU network to obtain the hidden layer state after time step weighting, and finally is processed through the fully connected layer to output the industrial load situation prediction result.

[0018] In a second aspect, the present invention provides an industrial load portrait and situation prediction system based on multi-time scale features, which adopts the above method and includes: an information collection module, which is used to collect electricity consumption information and equipment characteristics of industrial park users; an industrial load portrait module, which is used to introduce a graph attention network to construct an industrial load portrait model based on the collected electricity consumption information and equipment characteristics of industrial park users, and obtain industrial load portrait characteristics; a multi-time scale sequence feature input module, which is used to construct a multi-time scale sequence feature input model based on the collected electricity consumption information and equipment characteristics of industrial park users, and obtain multi-time scale sequence characteristics; a feature extraction module, which is used to construct a feature extraction model based on a feature multi-head self-attention layer, integrate the industrial load portrait characteristics with the multi-time scale sequence characteristics through the feature extraction model, use attention to weight each feature, quantify the correlation of each feature and focus on important features; an industrial load situation prediction module, which is used to construct a prediction model based on a time series multi-head self-attention layer, input the data output by the feature extraction model into the prediction model for training, and the trained prediction model is used to output the industrial load situation prediction result.

[0019] Compared with the prior art, the above technical solution has the following advantages: The present invention collects the electricity consumption information and equipment characteristics of industrial park users to realize the collection of diversified industrial loads, and establishes industrial load portraits based on this. It uses multi-time scale sequence features to fuse industrial load portrait features and combines a double-layer multi-head self-attention mechanism (feature multi-head self-attention layer and time series multi-head self-attention layer) to analyze diversified industrial loads. The feature multi-head self-attention layer uses attention weights to focus on important features and mine the correlation of input feature sequences. The time series multi-head self-attention layer captures the correlation of time series, strengthens the information expression of important feature variables and key time points, and mines the periodic trend characteristics and local fluctuation characteristics of industrial loads. At the same time, an industrial load situation prediction model is established, which can realize multi-time scale prediction of the energy consumption situation of industrial park loads. The prediction model has good applicability and can be actually applied to different industrial load scenarios. It solves the problem that the current situation prediction research is difficult to take into account the randomness of industrial load fluctuations and the regularity of industrial users' electricity consumption behavior. It significantly increases the flexible regulation space of industrial parks and improves the digital perception and flexible regulation level of demand side. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0021] Figure 1 A module block diagram of a multi-time-scale characteristic industrial load profiling and situation prediction system provided by a specific embodiment of the present invention; Figure 2 A flowchart of a method for industrial load profiling and situation prediction based on multi-time scale characteristics provided by a specific embodiment of the present invention; Figure 3 A flowchart of an industrial load profiling and situation prediction method based on multi-time scale characteristics is provided as another specific embodiment of the present invention.

[0022] The accompanying drawings are marked as follows: information collection module 1, industrial load profiling module 2, multi-time scale sequence feature input module 3, feature extraction module 4, industrial load situation prediction module 5. DETAILED DESCRIPTION

[0023] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0024] Please refer to Figure 1 The present embodiment provides an industrial load profiling and situation prediction system based on multi-time scale features, which mainly includes the following modules: an information acquisition module 1, an industrial load profiling module 2, a multi-time scale sequence feature input module 3, a feature extraction module 4 and an industrial load situation prediction module 5, which are respectively used to execute steps S1 to S5 of the industrial load profiling and situation prediction method based on multi-time scale features described below in the present embodiment.

[0025] Please refer to Figure 2 This embodiment provides a method for industrial load profiling and situation prediction based on multi-time scale characteristics, which is mainly implemented through the following steps: S1: Collect electricity usage information and equipment characteristics of users in industrial parks.

[0026] S2: Based on the collected electricity consumption information and equipment characteristics of industrial park users, the graph attention network is introduced to construct an industrial load portrait model (also known as: load cluster energy consumption portrait model), obtain industrial load portrait characteristics, use the load behavior association graph to characterize the association relationship between different load devices of users, and use the graph attention network to optimize the weight of the correlation between loads. The graph attention network (GAT) evolved from the graph convolutional network. By introducing the attention mechanism between graph nodes to assign different weights to their transmission parameters, it further characterizes the mutual correlation between nodes. GAT has the advantages of high precision and strong generalization ability. It solves the problem that the current load behavior association graph can only assign the same weight to the neighboring nodes, so as to pay more attention to the correlation characteristics between load nodes, so as to help more accurately capture the complex relationship between load behaviors, enhance the model's attention to the correlation between users' electricity consumption behaviors, and improve the accuracy of load prediction. And the information obtained is passed to the multi-time scale sequence feature input model.

[0027] S3: Based on the collected electricity consumption information and equipment characteristics of industrial park users, a multi-time scale sequence feature input model is constructed to obtain multi-time scale sequence features. Since the time series signal of industrial load changes complexly and presents different rules at different time scales, multi-time scale feature fusion is used to mine the periodic trend characteristics and local fluctuation characteristics of industrial load, improve the interpretability of industrial load representation, and pass the model data to the feature extraction model.

[0028] S4: Construct a feature extraction model based on the feature multi-head self-attention layer, integrate the industrial load portrait features with the multi-time scale sequence features through the feature extraction model, use attention to weight each feature, quantify the correlation of each feature and focus on important features. In this step, the received load cluster energy consumption portrait model is integrated with the multi-time scale sequence feature input model, and the feature multi-head self-attention layer in the double-layer multi-head self-attention layer uses attention weights to focus on important features. According to the actual interaction situation, the input feature sequence feature correlation model is constructed, in order to obtain an industrial load portrait and situation prediction system based on multi-time scale features.

[0029] S5: Build a prediction model based on the time series multi-head self-attention layer, input the data output by the feature extraction model into the prediction model for training, and the trained prediction model is used to output the industrial load situation prediction results. This step uses the BiGRU neural network regression prediction algorithm to train the input data through the data passed by the feature extraction model. The time series multi-head self-attention layer is applied to the output of the BiGRU network. By quantizing the weights, the time series multi-head self-attention layer can perceive key historical information, and the fully connected layer outputs the industrial load prediction results.

[0030] Embodiment 2

[0031] Please refer to Figure 3 This embodiment provides a more preferred system and method for industrial load profiling and situation prediction based on multi-time scale characteristics, which is mainly implemented through the following steps: S1: Collect electricity usage information and equipment characteristics of users in industrial parks.

[0032] S2: According to the collected electricity consumption information and equipment characteristics of industrial park users, a graph attention network is introduced to construct an industrial load portrait model to obtain industrial load portrait characteristics. This embodiment uses a load behavior association graph to characterize the association relationship between different load devices of users, and uses a graph attention network to optimize the weight of the correlation between loads, that is, the improved load behavior association graph extraction model based on the graph attention network constructs a load feature matrix for discrete and continuous feature data, and combines the graph attention network to construct an improved load behavior association graph. The power consumption in the power consumption information is an important manifestation of the load behavior. From the time dimension, the total power consumption within a period of time can reflect the overall scale of the load behavior. Moreover, the different usage time of different load devices will cause the load behavior to present different peaks and troughs. When multiple devices are used at the same time, the load behavior will show a superposition effect.

[0033] In S2, a graph attention network is introduced to construct an industrial load profiling model, which mainly includes the construction of load prior behavior association graph and the extraction of improved load behavior association graph based on graph attention network, as follows: S21: Construct a load prior behavior association graph based on formula (1), (2) and (3). According to the collected electricity consumption information and equipment characteristics of industrial park users, obtain the load node set of various load behaviors and load numbers of industrial park users and the directed edge set of transfer relations and association numbers between different load behaviors, construct a load prior behavior association graph represented by a directed graph, calculate the prior transfer probability between any load behaviors in the load prior behavior association graph, and construct the load prior behavior transfer matrix based on the prior transfer probability. More detailed steps are as follows: Using directed graph To represent the load prior behavior association graph, where V represents various load behaviors of industrial users and the node set with load number |V|; A directed edge set with an association number of |E| to characterize the transfer relationship between different load behaviors of users. Load prior behavior can provide guidance for predicting and understanding load behavior. Actual data of load behavior can verify the accuracy of load prior behavior. To construct a load prior behavior association graph, it is first necessary to calculate the prior transfer probability between any load behaviors (also known as: the directed edge weight of the load prior behavior association graph): (1) In the formula, Load node Runtime Load Node The probability of running, i.e. the prior transition probability; Load node The total number of prior occurrences; Load node With Node The total number of times the loads appear at the same time. The priori transition probability between load behaviors can help us understand the collaborative working mode between different loads and the user's behavior habits. For example, in a home's power system, it is the probability that the TV is turned on when the air conditioner is turned on.

[0034] To avoid overfitting of the correlation matrix, a threshold can be set Perform binary smoothing on the prior transfer probability and obtain , thus constructing the prior load behavior transfer matrix : (2) (3) In the formula, It is the proportion of historical load behaviors retained each time the user load prior behavior association diagram is updated. In addition, the association matrix here refers to a mathematical structure used to characterize the mutual correlation between different load behaviors of industrial users, which is constructed by calculating the prior transition probability between any load behaviors.

[0035] At this point, the user load prior behavior association graph is constructed.

[0036] S22: Extraction of improved load behavior association graph of graph attention network based on equations (4), (5) and (6). A shared weight parameter matrix is ​​introduced to transform the input feature of the feature vector of the load node in the load prior behavior transfer matrix as the input feature vector of the graph attention network. The output of the graph attention network is the attention coefficient of each adjacent node. The feature vector of the corresponding load node in the load prior behavior association graph in S21 is optimized by the attention coefficient to realize the extraction of improved load behavior association graph of graph attention network. The more detailed steps are as follows: Existing research shows that the load behavior of industrial users is often associated with the behavior of their adjacent loads and the factors affecting the load behavior. It can reflect the correlation strength of the directional transfer between the load behaviors at both ends of the edge (i.e., two adjacent load behaviors), but this is based on the statistical analysis of prior events and cannot characterize the impact of other influencing factors on the load behavior transfer. Therefore, the introduction of GAT helps to better obtain the correlation characteristics between different loads and their influencing factors.

[0037] For any time slot t, the load transfer matrix As the input feature vector of GAT, a shared weight parameter matrix W is set to load nodes through the shared weight parameter matrix W With adjacent load nodes The input feature is transformed, where the number of rows M in W is the number of output features. Then each of them passes through a single hidden layer neural network that represents the node attention mechanism. And Softmax activation function to calculate the attention coefficient between two nodes and normalize the attention coefficient: (4) (5) In the formula, and They are the load transfer matrices in time slot t respectively. The i-th and j-th columns, i.e., the load nodes and The input feature vector of A single hidden layer neural network is a parameterized weight vector of ; ‖ represents the concatenation of two vectors; LeakyReLU(·) is the activation function.

[0038] At any time slot t, the attention coefficient and its normalized value Characterizes the load node for For any load node , using the normalized attention coefficients of its neighboring nodes, the corresponding load feature vectors can be weighted summed as the output feature vector of the node after GAT .

[0039] (6) In the formula, (·) is a nonlinear activation function.

[0040] For any time slot t, the load transfer matrix is ​​used as the input feature vector, and the attention coefficients are obtained after a series of operations. These coefficients represent the load node for Then, in the load prior behavior association graph constructed by S21, the weight of its directed edge is originally constructed based on the prior transfer probability. When the attention coefficient is applied, the normalized attention coefficient of each adjacent node is used to optimize the feature vector of the corresponding load node, that is, the weighted sum of the corresponding load feature vector is obtained to obtain the new output feature vector The input of GAT is the feature vectors of multiple nodes. The weights between nodes are calculated through the attention mechanism, and then the weighted sum of the features of neighboring nodes is used to obtain the updated node feature representation. The LeakyReLU activation function is used when calculating the attention coefficient to introduce nonlinear characteristics.

[0041] The graph attention network helps to aggregate the correlation information between load events and neighbor events by mapping the load characteristics containing key factors such as weather and historical load into the correlation attention between different load behaviors.

[0042] S3: Based on the collected electricity consumption information and equipment characteristics of industrial park users, a multi-time scale sequence feature input model is constructed to obtain multi-time scale sequence features. This embodiment uses multi-time scale feature fusion to mine the periodic trend characteristics and local fluctuation characteristics of industrial loads, and passes the output data to the feature extraction model. The more detailed steps are as follows: S31: Process time series data of multiple time scales based on formula (7). It can be data of different frequencies (such as seconds, minutes, hours, days, etc.). Decompose the time series data into different time scales. For example, wavelet transform, Fourier transform and other techniques can be used to decompose the data into low-frequency and high-frequency components. Extract features from the decomposition results of different time scales. Fourier transform is shown in formula (7): (7) Where f(t) is a function in the time domain, which means a function of time t; F(w) is a function in the frequency domain, which means a function of frequency w; is a complex exponential function, where j is the imaginary unit and w is the angular frequency.

[0043] In view of the differences in the characteristics of industrial loads at multiple time scales, active power sequences of different time scales are included in the input of the neural network to achieve comprehensive learning of local fluctuation details and overall operating trends. For example, assuming that the input time intervals are 30s, 1min, and 4min, the active power time series, since the lengths are all 360, the input contains 3h, 6h, and 24h of information respectively. By combining multi-time scale power sequences, the input data contains both accurate fluctuation information near the prediction point and long-term industrial load operation status information.

[0044] S4: A feature extraction model is constructed based on the feature multi-head self-attention layer. The industrial load portrait features and multi-time scale sequence features are integrated through the feature extraction model. Attention is used to weight each feature, quantify the correlation of each feature and focus on important features. Specifically, it includes the following: S41: Based on formula (8), the weighted quantized feature matrix is ​​obtained. In the feature multi-head self-attention layer, the input feature sequence matrix is ​​weighted to obtain the weight of any feature sequence, thereby obtaining the weighted quantized feature matrix. The more detailed steps are as follows: The multi-time scale sequence feature data and the feature vector output by GAT are used as the feature sequence matrix to input into the feature multi-head self-attention layer. In the feature multi-head self-attention layer, the input feature sequence matrix X is weighted to quantify the correlation between each feature and its impact on the active power of industrial load. For any feature sequence i, the weight The calculation of is shown in formula (8): (8) Where X is the input feature sequence matrix.

[0045] S42: Based on formula (9) and (10), important features are captured at multiple scales and angles. The n query matrices are obtained from the input feature sequence matrix through the feature multi-head self-attention layer. After multi-head self-attention weighted quantization, the weighted matrix is ​​obtained, thereby realizing multi-scale and multi-angle capture of important features and multi-scale and multi-angle capture of feature importance. The more detailed steps are as follows: The Query matrix is ​​transformed linearly 、Key Matrix and Value matrix Projected into 3 independent spaces; , and is the corresponding transformation matrix; (·) is the multi-head self-attention output function, and its calculation process is shown in formula (9): (9) In the formula, is the attention score matrix, T represents transpose; is the dimension of the Key matrix; is the i-th self-attention output; Concat(·) is the concatenation function; To convert the matrix, the concatenated matrix can be converted to a specific length. The multi-head self-attention mechanism is achieved by Project X to different subspaces to obtain n query matrices, realizing multi-scale and multi-angle capture of feature importance. According to the attention weights calculated by equations (8) and (9), the weighting matrix is ​​obtained by equation (10): : (10) In the formula, X is the input feature sequence matrix, Refers to the multi-head self-attention output function. After X is optimized by the feature multi-head self-attention layer, the correlation between features is effectively reflected in the weight of each feature. The feature multi-head self-attention layer enhances the key influencing factors of the model on the power output of industrial loads, suppresses the influence of weakly correlated features, and improves the prediction accuracy by perceiving the contribution of features to the results.

[0046] In the method proposed in the present invention, multiple self-attention heads can respectively learn different representations of MTS active power, current and other features, so as to have a deeper interpretation of power mutation, change trend and feature correlation in the prediction task. Among the different features of the model input, the power factor can reflect the changes in the load resistance, inductance and capacitance, and its trend is associated with the change trend of industrial loads of different categories. Different load current change patterns are different, and the load prediction results can be improved by extracting features. For example, motor-type industrial equipment usually requires a larger starting current during the startup phase, while pure resistance equipment has a weaker starting current requirement. Therefore, when facing different industrial equipment, different electrical or time features have different degrees of influence on the load prediction task. The multi-head self-attention mechanism is a core component in the Transformer architecture. By projecting the input features into multiple different subspaces, the feature representations of different aspects are learned respectively, so as to achieve multi-scale and multi-angle capture of the importance of features.

[0047] S5: Construct a prediction model based on the time series multi-head self-attention layer, input the data output by the feature extraction model into the prediction model for training, and the trained prediction model is used to output the industrial load situation prediction results. In this step, the input data is trained using the BiGRU neural network regression prediction algorithm to achieve the final load prediction value. Among them, the BiGRU network includes an input layer, a hidden layer, and an output layer. At any moment, the hidden layer state output by the BiGRU neural network is jointly determined by the output results of the forward GRU and the backward GRU at that moment. The weight matrix obtained in S42 is input into the BiGRU neural network for feature training to obtain the updated hidden layer state; the time series multi-head self-attention layer is applied to the output end of the BiGRU network. The time series multi-head self-attention layer quantizes the weights of the hidden layer state output by the BiGRU network to obtain the hidden layer state after time step weighting, and finally processes it through the fully connected layer to output the industrial load situation prediction results. More detailed steps are as follows: S51: Feature learning is performed at the BiGRU layer based on equations (11) and (12). At any moment, the hidden layer output (total output value) of the BiGRU is jointly determined by the output results of the forward GRU and the backward GRU at that moment, as shown in equation (11): (11) In the formula, and Respectively represent the hidden layer states of the forward GRU and backward GRU output at time t; is the hidden layer state of BiGRU output at time t; and are the weights corresponding to the hidden layer states of the forward GRU and backward GRU respectively; is the corresponding bias coefficient.

[0048] enter To the BiGRU layer for feature learning, the hidden layer state is updated as shown in formula (12): (12) Where f is the BiGRU network structure. Refers to the hidden layer output of the forward GRU at time t, is the weight matrix, is the weight matrix at time t.

[0049] S52: The prediction result after time step weighting is obtained based on formula (13). The operation of industrial equipment is often based on the production cycle plan, and its load has certain regularity and continuity. Therefore, the industrial load is similar to its historical operation, and different time steps have different effects on the prediction results. In order to adaptively analyze the cross-correlation of time steps and obtain the weight of the hidden state of the time series, the time series multi-head self-attention layer is applied to the output of the BiGRU network. By quantizing the weights, the time series multi-head self-attention layer can perceive key historical information, and its calculation process is shown in formula (13): (13) Where H is the hidden layer state output by the BiGRU layer, , and is the weight matrix of the i-th head in the temporal multi-head self-attention layer. is the hidden layer state after weighting the time step, (·) is the multi-head self-attention output function.

[0050] S53: The fully connected layer outputs the industrial load forecast result Y. Y can predict the industrial load situation within the next hour (i.e. 120 time steps). The fully connected layer is located in the last stage of the model. Its main function is to receive the output from the time series multi-head self-attention layer and convert it into the final industrial load forecast result. It will perform a linear transformation on the input features, and integrate and map the information extracted and processed by the previous layers through the learned weights and bias parameters.

[0051] During the training process, the fully connected layer will continuously adjust its weights and biases according to the training data to minimize the error between the predicted result and the actual value (such as the root mean square error RMSE and the mean absolute percentage error MAPE). In this way, the model can output industrial load forecast values ​​that are consistent with the actual situation.

[0052] Through the above embodiments, the industrial load profiling and situation prediction system and method based on multi-time scale features of the present invention can achieve the following technical effects compared with the prior art: (1) The load behavior association graph is used to characterize the association relationship between different load devices of the user, and the graph attention network is used to optimize the weight of the correlation between loads, which can more accurately reflect the association relationship between load devices. (2) The cyclic trend characteristics and local fluctuation characteristics of industrial loads are mined by multi-time scale feature fusion to improve the interpretability of industrial load characterization. (3) The dual-layer multi-head self-attention (DMHSA) mechanism uses attention weights to focus on important features, captures time series correlation while mining the correlation of input features, and strengthens the information expression of important feature variables and key time steps.

[0053] In summary, the present invention collects the electricity consumption information and equipment characteristics of industrial park users to realize the collection of diversified industrial loads, and establishes industrial load portraits based on this. It uses multi-time scale sequence features to fuse industrial load portrait features and combines a double-layer multi-head self-attention mechanism (feature multi-head self-attention layer and time series multi-head self-attention layer) to analyze diversified industrial loads. The feature multi-head self-attention layer uses attention weights to focus on important features and mine the correlation of input feature sequences. The time series multi-head self-attention layer captures the correlation of time series, strengthens the information expression of important feature variables and key time points, and mines the periodic trend characteristics and local fluctuation characteristics of industrial loads. At the same time, an industrial load situation prediction model is established, which can realize multi-time scale prediction of the energy consumption situation of industrial park loads. The prediction model has good applicability and can be actually applied to different industrial load scenarios. It solves the problem that the current situation prediction research is difficult to take into account the randomness of industrial load fluctuations and the regularity of industrial users' electricity consumption behavior. It significantly increases the flexible regulation space of industrial parks and improves the digital perception and flexible regulation level of demand side.

[0054] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the above-mentioned embodiments of the methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0055] The technical features of the above-described embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above-described embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification. The above-described embodiments only express several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they cannot be understood as limiting the scope of the present invention. It should be pointed out that for those of ordinary skill in the art, without departing from the concept of the present invention, several variations and improvements can be made, which all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.

Claims

1. A method for industrial load profiling and situation prediction based on multi-time scale characteristics, characterized in that: These include: S1: Collect electricity consumption information and equipment characteristics of users in industrial parks; S2: Based on the collected electricity consumption information and equipment characteristics of industrial park users, the graph attention network is introduced to build an industrial load portrait model to obtain industrial load portrait characteristics; S3: Based on the collected electricity consumption information and equipment characteristics of industrial park users, a multi-time scale sequence feature input model is constructed to obtain multi-time scale sequence features; S4: A feature extraction model is constructed based on the feature multi-head self-attention layer. The industrial load portrait features and multi-time scale sequence features are integrated through the feature extraction model. Attention is used to weight each feature, quantify the correlation of each feature and focus on important features. S5: A prediction model is constructed based on the time series multi-head self-attention layer. The data output by the feature extraction model is input into the prediction model for training. The trained prediction model is used to output the industrial load situation prediction results.

2. The method for industrial load profiling and situation prediction based on multi-time scale characteristics according to claim 1 is characterized in that: S2 specifically includes: S21: Construct a priori load behavior association graph: Based on the collected electricity consumption information and equipment characteristics of industrial park users, obtain the load node sets of various load behaviors and load numbers of industrial park users and the directed edge sets of transfer relations and association numbers between different load behaviors, construct a load priori behavior association graph represented by a directed graph, calculate the priori transfer probability between any load behaviors in the load priori behavior association graph, and construct a load priori behavior transfer matrix based on the priori transfer probability; S22: Improved load behavior association graph extraction of graph attention network: Introduce a shared weight parameter matrix to perform input feature transformation on the feature vector of the load node in the load prior behavior transfer matrix as the input feature vector of the graph attention network. The graph attention network outputs the attention coefficient of each adjacent node. The feature vector of the corresponding load node in the load prior behavior association graph in S21 is optimized through the attention coefficient to achieve improved load behavior association graph extraction of the graph attention network.

3. The method for industrial load profiling and situation prediction based on multi-time scale characteristics according to claim 2 is characterized in that: In S21, binary smoothing is performed on the calculated priori transfer probabilities between any load behaviors, and the load priori behavior transfer matrix is ​​constructed according to the priori transfer probabilities after the binary smoothing.

4. The method for industrial load profiling and situation prediction based on multi-time scale characteristics according to claim 2 or 3 is characterized in that: S2 also includes: using the Softmax activation function to normalize the attention coefficients of each adjacent node, and optimizing the feature vector of the corresponding load node through the normalized attention coefficients.

5. The method for industrial load profiling and situation prediction based on multi-time scale characteristics according to claim 4 is characterized in that: The optimization method is: performing weighted summation on the feature vectors of the corresponding load nodes through the normalized attention coefficients of each adjacent node to obtain a new output feature vector of the corresponding load node.

6. The method for industrial load profiling and situation prediction based on multi-time scale characteristics according to claim 1 is characterized in that: In S3, the multi-time scale sequence feature input model is constructed using wavelet transform or Fourier transform, which is used to decompose time series data into sequence data of different time scales and extract features from the decomposition results of different time scales.

7. The method for industrial load profiling and situation prediction based on multi-time scale characteristics according to claim 1 is characterized in that: S4 specifically includes the following: S41: Obtaining a weighted quantized feature matrix: weighting the input feature sequence matrix in the feature multi-head self-attention layer to obtain the weight of any feature sequence, thereby obtaining a weighted quantized feature matrix; S42: Capturing important features at multiple scales and angles: Through the feature multi-head self-attention layer, n query matrices are obtained from the input feature sequence matrix. After multi-head self-attention weighted quantization, the weighted matrix is ​​obtained, thereby realizing multi-scale and multi-angle capture of important features.

8. The method for industrial load profiling and situation prediction based on multi-time scale characteristics according to claim 7 is characterized in that: In S5, the prediction model uses a BiGRU neural network to train the input data; the BiGRU network includes an input layer, a hidden layer and an output layer.

9. The method for industrial load profiling and situation prediction based on multi-time scale characteristics according to claim 8 is characterized in that: At any moment, the hidden layer state output by the BiGRU neural network is jointly determined by the output results of the forward GRU and the backward GRU at that moment. The weighting matrix obtained by S42 is input into the BiGRU neural network for feature training to obtain an updated hidden layer state; the temporal multi-head self-attention layer is applied to the output end of the BiGRU network, and the temporal multi-head self-attention layer quantizes the weights of the hidden layer state output by the BiGRU network to obtain the hidden layer state after time step weighting, and finally is processed by the fully connected layer to output the industrial load situation prediction result.

10. A system for industrial load profiling and situation prediction based on multi-time scale characteristics, using the method described in any one of claims 1 to 9 above, characterized in that: include: Information collection module, used to collect electricity consumption information and equipment characteristics of users in industrial parks; The industrial load portrait module is used to introduce the graph attention network to build an industrial load portrait model based on the collected power consumption information and equipment characteristics of industrial park users, and obtain the industrial load portrait characteristics; The multi-time scale sequence feature input module is used to construct a multi-time scale sequence feature input model based on the collected electricity consumption information and equipment characteristics of industrial park users to obtain multi-time scale sequence features; The feature extraction module is used to build a feature extraction model based on the feature multi-head self-attention layer. The feature extraction model integrates the industrial load portrait features and multi-time scale sequence features, uses attention to weight each feature, quantifies the correlation of each feature, and focuses on important features. The industrial load situation prediction module is used to build a prediction model based on the time series multi-head self-attention layer. The data output by the feature extraction model is input into the prediction model for training. The trained prediction model is used to output the industrial load situation prediction results.

Images