Load demand response characteristic analysis and prediction method based on load side data mining

Abstract

The invention discloses a load demand response characteristic analysis and prediction method based on demand side data mining, and provides an improved fuzzy C-means clustering method suitable for multi-source massive heterogeneous demand side power data. Dimensionality reduction preprocessing is carried out on user-side power data, data features can be efficiently extracted while complexity is reduced, the difficulty of constructing an adjustable resource pool with a construction load is reduced, and analysis of a power grid company on user consumption behaviors is facilitated; a multi-type load prediction model based on a convolutional neural network is provided. And selecting high-correlation data by using CNN, and screening an input variable set of the high-correlation input prediction model. Establishing a load prediction model based on Transform, and applying the load prediction model to load prediction of a demand side adjustable resource pool; the power utilization characteristics of multiple types of loads and various factors influencing participation in demand response are analyzed in detail, the response capability and characteristics of the adjustable resources on the demand side are determined and quantified in combination with a load prediction data set, the multi-dimensional characteristics of the adjustable resources on the load side are mined, and the potential of the adjustable resources on the demand side in supply and demand adjustment is exerted. And the interactive operation characteristics and the power consumption condition of the demand side resources are mastered.

Description

Technical Field

[0001] The present invention relates to the field of demand-side technology, and in particular to a method for analyzing and predicting demand-side load demand response data mining characteristics. Background Art

[0002] Load demand response characteristic analysis based on demand-side data mining is a key system performance optimization method. Its goal is to deeply understand the demand response characteristics of the system under different load conditions by mining the data of system users or business parties, so as to better meet user expectations, improve system performance, and ensure the stability of the system under various workloads.

[0003] Building a demand-side resource library requires integrating massive amounts of distributed load data, making it difficult to segment and aggregate massive, multi-source, heterogeneous user-side adjustable resources. With the large-scale construction and typical application of adjustable resource pools, existing and incremental data are characterized by large quantities, strong redundancy, and high computational complexity. This imbalance in load power consumption and diverse types results in the grid's inability to fully mobilize demand-side resources for demand response, resulting in wasted adjustable resources. Therefore, it is necessary to research new analytical techniques to mine massive amounts of user data to accurately quantify and efficiently utilize demand-side resources. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] To solve the above technical problems, the present invention provides the following technical solutions: a method for analyzing and predicting the characteristics of demand-side load demand response data mining, comprising:

[0006] An improved fuzzy C-means clustering method is proposed for large amounts of heterogeneous demand-side power data from multiple sources. Dimensionality reduction preprocessing of user-side power data reduces complexity while effectively extracting data features, making it easier to construct an adjustable resource pool under load conditions and facilitating the analysis of user consumption behavior by power grid companies. A multi-type load forecasting model based on a convolutional neural network is proposed. The CNN is used to select highly correlated data and filter the set of highly correlated input variables for the forecasting model. A transformer-based load forecasting model is established and applied to load forecasting for a demand-side adjustable resource pool. The power consumption characteristics of various load types and the factors influencing demand response participation are analyzed in detail. Combined with the load forecasting dataset, the response capabilities and characteristics of demand-side adjustable resources are determined and quantified. This approach explores the multidimensional characteristics of load-side adjustable resources and leverages their potential in supply and demand regulation. This approach also captures the interactive operation characteristics and power consumption of demand-side resources.

[0007] A method for analyzing and predicting load demand response characteristics based on demand-side data mining is proposed. By mining data on demand-side adjustable resources and analyzing the response characteristics of various load types, combined with a predictive model, accurate prediction and measurement of demand-side load flexibility can be achieved. This effectively guides various load types to participate in grid interaction in a targeted manner, improving the system's frequency and peak-shaving capabilities, and ultimately achieving comprehensive optimization of power consumption and system resources. This method can effectively assist in the construction of a large-capacity adjustable load resource library, playing a vital role in power demand response, ensuring supply and demand balance, promoting new energy consumption, and ensuring the stability and reliability of power supply.

[0008] As a preferred method for mining demand-side adjustable resource data according to the present invention, a method for preprocessing power load sample data features based on improved fuzzy C-means clustering is adopted, including:

[0009] The complexity of power load data lies in the fact that there are many factors that affect the load, which leads to its high uncertainty. Therefore, the preprocessing of load data is an indispensable operation before prediction. Using clustering algorithms to cluster load data and analyze load types is a common and effective method to reduce the complexity of load data. However, due to the large scale and complexity of massive multivariate heterogeneous data on the load side, traditional clustering algorithms such as C-means and fuzzy C-means that are commonly used for power load data preprocessing have the problem of local convergence caused by improper selection of initial cluster centers, which affects the effect of power load preprocessing and further affects the accuracy of load prediction. To this end, the present invention proposes a method for preprocessing the characteristics of power load sample data based on an improved fuzzy C-means clustering algorithm. By improving the selection of cluster centers by the traditional fuzzy C-means algorithm, the power load type identification effect is improved, laying the foundation for subsequent power load prediction.

[0010] Data preprocessing includes normalization and data partitioning. Dataset of historical data for days, ,in Indicates the The dimension m of the load vector for each day depends on the sampling frequency of the data set.

[0011] First, the historical load data set is normalized on a daily basis, and all load values ​​in a day are divided by the maximum load value of that day: #timg# （1）

[0012] Where, is the normalized load vector, For the The load vector of the day, max() is the maximum value function.

[0013] Secondly, the normalized data set is randomly divided into two parts, namely the training set and the test set, respectively. and express, and Represents the number of days included in the training set and test set respectively. The training set contains load data for 80% of the days, and the test set contains load data for the remaining 20% ​​of the days.

[0014] As a preferred method for mining demand-side adjustable resource data in the present invention, a fuzzy C-means clustering algorithm model based on SSA optimization is adopted. The algorithm flow is as follows: Figure 1 shown.

[0015] To improve the clustering performance of the traditional fuzzy C-means clustering algorithm for large-scale, multivariate, and heterogeneous data on the load side, an improved fuzzy C-means clustering algorithm is proposed, using an auxiliary neural network to extract the mapping relationship between external factors such as date and weather and node daily loads. This algorithm is then used for power load classification and identification. Taking into account the significant differences between different load types, an improved fuzzy C-means clustering algorithm based on sparrow optimization is proposed. Load samples are classified according to climate and date characteristics, and a discriminant interval between the Euclidean distance between load characteristic indicators and the fuzzy C-means is constructed. In the clustering results, the load curve with the greatest distance from the remaining categories is considered an atypical daily load curve. By eliminating atypical daily load curves with unclear classification characteristics and screening out categories with low correlation, data dimensionality reduction is achieved.

[0016] The fuzzy C-means clustering method is an unsupervised clustering algorithm that uses Euclidean distance for discriminant analysis. In the initial iteration, the algorithm randomly generates cluster centers. By measuring the Euclidean distance between samples, samples with similar distances are grouped together. The cluster centers are updated by calculating the mean of each class of samples, and the next iteration begins. The calculation process for updating cluster centers and achieving clustering using the fuzzy C-means clustering method is shown below.

[0017] Assume that after normalization data samples ,in It is A load vector. is the number of categories into which the data samples are to be divided, Indicates the corresponding categories, is its similarity classification matrix, and the cluster center points of each category are , It is a sample for The membership degree (abbreviated as ). Then the objective function It can be expressed as follows: #timg# （2）

[0018] Where, . is the Euclidean distance, which is used to measure the samples With the The distance between cluster centers; is the number of features of the sample; is a weighting parameter, and its value range is The fuzzy C-means algorithm is to find an optimal classification so that the classification can produce the minimum function value. It requires that the sum of the membership values ​​of a sample to each cluster is 1, that is, it satisfies: #timg# （3） #timg# （4）

[0019] Formula (3) and formula (4) are used to calculate the sample For the membership of A1 and c cluster centers .

[0020] set up , for all kind, , =0. #timg# （5）

[0021] Since the fuzzy C-means clustering method randomly generates cluster centers in the initial iteration, the clustering may fall into a local optimal solution. To address this problem, this study uses the SSA algorithm to optimize the initial cluster centers of the fuzzy C-means clustering method. SSA is an optimization algorithm that simulates the search behavior of sparrows. The sparrows in the population are divided into discoverers, joiners, and guards. The ratio of discoverers to joiners remains unchanged, but their identities change dynamically. Each sparrow is randomly given an initial position and a fitness that represents energy reserves, which are updated in each iteration. The formula for updating the discoverer's position during the iteration process is as follows: #timg# （6）

[0022] Where, Indicates Time Only sparrow Dimensional position, and for (0,1] A random number uniformly distributed in the interval, is a standard normally distributed random number, Indicates the warning threshold, and its value range is [0.5,1.0] , is the maximum number of iterations. From the formula, we can get: and The relative size of determines the way the discoverer updates its position. In SSA, the joiner's position update follows the following formula: #timg# （7） #timg# （8）

[0023] Where, It is composed of 1 or -1, Matrix, where Indicates the variable dimension to be optimized, is a matrix The value in dimension d, represents the position of the current global worst finder, is the best position among the current finders. Indicates the joiner, Indicates the number of sparrow species. The location of the joiner is updated by and Determined by the relative size of.

[0024] In addition, there are a certain number of watchers in the SSA, accounting for 10% to 20% of the total number of sparrows. When there is a danger, the discoverer or joiner will be updated according to the following rules. #timg# （9）

[0025] Where, is the parameter representing the step size control, The value is [-1,1] A random number, represents the globally optimal and safe sparrow position. , and Represent the fitness values ​​of the current sparrow, the current best sparrow and the current worst sparrow respectively. Set a small constant to avoid the denominator being 0.

[0026] Discoverers are primarily responsible for global search, while joiners join the optimal solution through local search. Watchers are responsible for preventing the algorithm from falling into local optimal solutions. The ratio of discoverers to joiners remains constant, while their identities change dynamically, and their positions are updated in each iteration to approach the optimal solution. This optimization method, which combines global exploration and local search strategies, effectively avoids falling into local optimal solutions and achieves rapid convergence, resulting in high efficiency and accuracy in solving complex optimization problems.

[0027] Based on the global optimization capability of the SSA optimization algorithm, the position coordinates of the initial cluster center are defined as the sparrow position and the appropriate number of iterations is configured for continuous search to obtain the optimal initial cluster center, which is then used as the initial cluster center for calculation of the fuzzy C-means algorithm. By improving the cluster center of the fuzzy C-means, its clustering accuracy is improved. In order to more accurately identify the categories after clustering, a small number of edge samples in a certain category have unclear similarities in their features and can be determined to be outlier load samples. Using this type of data does not utilize the characteristics of the neural network to learn the data. Therefore, the Euclidean distance threshold interval is further performed on the clustered curve. The data in this interval is identified as having unclear features and belonging to the outlier data. Its regular characteristics are not obvious, which is not conducive to neural network training. Such data is classified as abnormal data and directly solved using the traditional unit combination method.

[0028] As a preferred method for demand-side adjustable resource forecasting described in the present invention, a multi-type load demand forecasting model method based on CNN-Transformer is adopted, including:

[0029] Because power load data does not have actual category "labels," it is impossible to apply metrics commonly used to evaluate clustering algorithms, such as accuracy and recall, and thus it is impossible to directly evaluate the effectiveness of power load clustering. The original intention of power load clustering is to reduce the complexity of load data and thereby improve the accuracy of power load forecasting. Therefore, based on the power load clustering results in the previous section, this study designed a CNN-Transformer-based power load forecasting model to verify the effectiveness of power load clustering and also to predict the load of demand-side adjustable resources. The CNN-Transformer model uses the CNN to capture the local features of the sequence and the Transformer model to capture the global dependencies between sequences. It can more effectively model the characteristics of the sequence and improve the performance of the model.

[0030] As a preferred method for forecasting demand-side adjustable resources described in the present invention, a time feature selection model method is adopted, including:

[0031] The CNN model is used to extract the time series features of factors that affect load, such as date and weather. The CNN model has two characteristics: one is that each layer in the CNN has a local connection with the next layer, and the other is that the weight value is shared throughout the CNN model. The characteristics of the CNN model enable it to perform high-dimensional processing on raw data and automatically and effectively extract the internal features of this data. The CNN model is as follows: Figure 2 The multi-layer combination shown in the figure has parameters such as the number of convolution layers and convolution kernels, which are referred to in the literature. The convolution layer and pooling layer are used to obtain information, which reduces the steps of feature extraction and the complexity of the operation, thereby improving the quality of the results.

[0032] Input layer: The number of units in the input layer is equal to the dimension of the predictor variable. The power load data needs to be preprocessed before being input into the model training.

[0033] Convolutional layer: The convolutional layer combines the correlation terms of the prediction indicators by adding a convolution operation between the input and output. The convolutional layer first linearly extracts information from the neurons in the input layer, then activates the neurons using a nonlinear activation function, and finally restores the activated neurons to an aggregate signal for output. A convolutional layer can be divided into three layers: input layer , convolution kernel (filter), convolution result layer Assumptions The sample matrix of an input signal of the layer is .filter The size of U V, then each neuron obtained in the L2 layer It's a signal and filter Convolution of different filters Different information of signal samples can be extracted. Here we assume that the output of convolution is Subscript from start: #timg# （10）

[0034] use Represents the convolution operation, the convolution layer uses the mapping of the previous layer With filter Perform convolution operation to extract information linearly. In fact, the filter It also has a bias. After the convolution operation, the bias constant is added to the convolution result. Get the convolution result signal That is The net input of the layer. Layer applies nonlinear activation function Activate the neurons and restore them to aggregate signals. Finally, linearly summarize the results of each feature map and input them into the next layer. The overall operation process of the convolutional layer is as follows: #timg# （11） #timg# （12）

[0035] Pooling layer: In a general CNN network, the pooling layer is located after the convolution layer. It uses the weight matrix of the sampling kernel to measure the overall characteristics of the data information, ignores unimportant subtle features, and achieves downsampling of the convolution layer neurons. Suppose the input matrix of the pooling layer is , the sampling kernel is essentially a weight matrix , the pooling layer is based on The input data is divided into pooling areas, and then the sampling kernel convolution pooling area is used to generate neurons in the pooling layer, that is, downsampling data The downsampling calculation formula is as follows: #timg# （13）

[0036] In the formula is the first downsampling layer elements, yes The corresponding pooling area.

[0037] Fully connected layer: The fully connected layer is usually located at the end of CNN. Its function is to aggregate all weighted signals into the final prediction, map the multi-dimensional vector into a one-dimensional vector, integrate all learned features and map them into the output space to give the classification result. The fully connected layer in the CNN network can be divided into Layer and layer. The function of the layer is to map the weighted signal obtained after processing by the convolution layer and pooling layer in the CNN network to the sample label space; The role of the layer is to use the activation function to obtain the final probability of classification prediction.

[0038] Assume that the input of the fully connected layer x=[ x 1 , x 2 ,…, x N 1 ] , yes The number of neurons in a layer is Layer and The layers are connected as follows: #timg# （14）

[0039] Where, For output, and They are The weights and biases of the layers, for Activation function of the layer

[0040] Output layer: In the CNN network, the layer below the fully connected layer is the output layer, which is used to finally output the feature vector extracted by the entire network.

[0041] Activation function: In order to enhance the learning ability of the network and reduce overfitting, an activation function is usually used to activate the neural network. 、 etc. are commonly used activation functions. As an activation function, it enhances the nonlinearity of the model and reduces the possibility of data divergence. Its function is defined as: #timg# （15）

[0042] The role of CNN is to extract feature information from the input sequence. The activation function is used to activate the one-dimensional convolution layer 1 and the one-dimensional convolution layer 2, and the maximum pooling is selected as the pooling method of the pooling layer 1 and the pooling layer 2. After the original data is processed by the convolution layer and the pooling layer, it will first be mapped to the feature space of the hidden layer, and then the extracted features will be output through the fully connected layer structure. The fully connected layer uses the activation function The output feature vector of the CNN layer is The CNN process is expressed as: #timg# （16） #timg# （17） #timg# （18） #timg# （19） #timg# （20）

[0043] Among them, the outputs of convolution layer 1, convolution layer 2 and pooling layer 1, pooling layer 2 are respectively 、 and 、 , 、 and is the weight matrix of each layer of CNN, 、 、 、 and is the deviation value, and max() are convolution operations and maximum functions, and the output length of the CNN layer is

[0044] As a preferred method for forecasting demand-side adjustable resources described in the present invention, a load demand forecasting method is adopted, including:

[0045] Based on CNN's extraction of time series features from raw data, a Transformer-based short-term power load forecasting method is proposed. The Transformer algorithm, a combination of fully connected layers and a self-attention mechanism, is a model structure that avoids meaningless loops and relies entirely on its attention mechanism to model the global dependencies between inputs and outputs. It primarily uses an encoder-decoder approach for time series forecasting, boasting excellent parallelism and effectively achieving optimal results.

[0046] The Transformer encoder captures long-term dependencies in the load sequence and learns and memorizes periodic patterns in the load data. The decoder receives the encoder output and generates a load demand forecast based on it. Using a self-attention mechanism and an attention mechanism on the encoder output, the decoder dynamically weights the input sequence, generates a contextual representation relevant to the prediction, and ultimately generates a prediction.

[0047] The self-attention mechanism is the core of the Transformer algorithm and can produce more interpretable models. The attention distribution can be examined within the model. Specifically, the input data is encoded as a position vector, which is then converted into the Q, K, and V vectors required by the self-attention mechanism using the three weight matrices WQ, WK, and WV. The self-attention value is then calculated.

[0048] The Transformer uses only one self-attention algorithm to predict load demand. Its multi-head attention mechanism repeats this algorithm multiple times, resulting in a more balanced and targeted result. During training and prediction, the Transformer effectively extracts intrinsic information from the data by learning and training on it. It also masks the original data during prediction training, making the algorithm's predictions more convincing.

[0049] The process of short-term load forecasting based on Transformer: The load data clustered by IFCM and other factors affecting the load that have been processed by CNN feature extraction are extracted through position encoding and then input into the Transformer encoder to capture the dependency of time series data. The prediction result is then obtained through the fully connected layer, such as Figure 3 shown.

[0050] The specific steps are as follows:

[0051] Input the correlation scores between similar daily load, historical load and multi-dimensional related influencing factors screened by CNN, and calculate the correlation scores using the dot product method, which is to use Each vector in Calculate the dot product for each vector in . The specific formula is as follows: #timg# （21）

[0052] Normalize the correlation scores between features in the input sequence to stabilize the gradient during training. The specific formula is as follows: #timg# （22）

[0053] Using the normalization function, the score vector between each input value is converted into [0,1] The probability distribution between them, and at the same time, it highlights the relationship between the input values. After normalization, the score is converted into a value distributed in [0,1] The matrix between .

[0054] According to the probability distribution between different types of load inputs, then multiply the corresponding value, and Perform the dot product.

[0055] Adjust appropriate parameters, perform training and prediction on the required data based on the obtained data, and test the prediction results.

[0056] Set the success rate threshold. When the threshold is reached Or when the maximum number of training times has been reached, it stops and outputs the prediction results.

[0057] The error analysis of the prediction results was performed, and the mean absolute error (MAE) and root mean square error (RMSE) were used as measurement indicators. The specific calculation formula is as follows: #timg# （23） #timg# （24） BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive work, among which:

[0059] Figure 1 Schematic diagram of the fuzzy C-means clustering process based on SSA optimization of the present invention;

[0060] Figure 2 This is a network structure diagram for CNN prediction in the present invention;

[0061] Figure 3 This is a flow chart of the Transformer-based load forecasting method of the present invention;

[0062] Figure 4 This is a comprehensive comparison chart of the performance of different prediction models based on the present invention. DETAILED DESCRIPTION

[0063] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0064] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0065] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0066] The present invention is described in detail with reference to schematic diagrams. For ease of illustration, cross-sectional views of device structures may be partially enlarged and not to scale when describing embodiments of the present invention. Furthermore, the schematic diagrams are merely illustrative and should not limit the scope of the present invention. Furthermore, in actual production, the three-dimensional dimensions of length, width, and depth should be included.

[0067] In the description of the present invention, it should be noted that the terms "upper, lower, inner, and outer" and other references to orientations or positional relationships are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate and simplify the description of the present invention. They are not intended to indicate or imply that the devices or components referred to must have, be constructed, or operate in a specific orientation, and therefore should not be construed as limitations on the present invention. Furthermore, the terms "first, second, or third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0068] In this disclosure, unless otherwise specified or limited, the terms "mounted," "connected," and "connected" should be interpreted broadly. For example, they may refer to fixed, removable, or integral connections. They may also refer to mechanical, electrical, or direct connections, indirect connections through an intermediary, or internal communication between two components. Those skilled in the art will understand the specific meanings of these terms in this disclosure.

[0069] Example 1

[0070] The first embodiment of the present invention provides a method for constructing a load characteristic index to extract effective features of the load curve, and assisting a clustering algorithm to classify the curve. Before clustering the load curve, the effective features of the load curve are extracted by constructing a load characteristic index, and the curve is assisted by a clustering algorithm. As a preferred method for analyzing the demand response characteristics of demand-side adjustable resources described in the present invention, a method for constructing a load characteristic index and a clustering index is adopted, including: before clustering the load curve, the effective features of the load curve are extracted by constructing a load characteristic index, and the curve is assisted by a clustering algorithm. This study is divided into four time period categories and 6 daily load characteristic indicators are selected. The time distinction is shown in the following table.

[0071] Load characteristic index time period classification Time period distinction Time Node All day 1:00-24:00 Peak 8:00-11:00，18:00-21:00 Flat 6:00-8:00，11:00-18:0021:00-22:00 Valley period 22:00-24:00，1:00-6:00

[0072] Select Reflects the average load of the whole day, load rate Reaction time utilization efficiency, daily peak-to-valley difference rate Reflects peak-to-valley differences. #timg# （25）

[0073] Where, 、 、 Respectively represent the total power, maximum power and minimum power in a day.

[0074] Peak load factor Reflects peak load changes. #timg# （26）

[0075] Where, The average power value for the whole day. Indicates the average value of the peak period in a day.

[0076] Average load factor Reflects the load changes during the normal period. #timg# （27）

[0077] Where, The load average value during the flat period.

[0078] Off-peak load rate Reflects the load changes during the valley period. #timg# （28）

[0079] Where, is the average load value during the valley period.

[0080] Clustering effectiveness indicators are usually used to evaluate the quality of clustering results or to select the appropriate number of clusters. This study selected two clustering indices, Davies-Bouldin Index (DBI) and Calinski-Harabaz Index (CHI), for quantitative analysis. Assuming that the dataset has The samples are clustered into kind.

[0081] The DBI index is called the classification accuracy index, which is used to measure the mean of the maximum similarity of each cluster and to measure the degree of closeness within a category and the degree of difference between categories [5]. #timg# （29） #timg# （30）

[0082] Where, Indicates the Class and The distance between cluster centers, is the maximum similarity value, and is the average distance from samples in each class to the center.

[0083] The CHI metric is calculated by comparing the separation index to the compactness index. A larger CHI indicates a tighter intra-class distribution and a more dispersed inter-class distribution, indicating a better clustering effect. #timg# （31）

[0084] Where, is the number of samples, is the number of categories, is the intra-class sum of squared errors, which is used to measure the closeness within the class. is the sum of squared errors between classes, which is used to measure the degree of separation between classes.

[0085] Example 2

[0086] The second embodiment of the present invention comprehensively evaluates the performance of the proposed model and compares it with various prediction models: TCN, LSTM, BiTCN, BiLSTM, CNN-GRU and LSTM-TCN.

[0087] Figure 4The error metrics for each model are illustrated. Compared to the LSTM-TCN model, the CNN-GRU demonstrates superior performance due to its bidirectional structure, which facilitates both forward and backward feature extraction. The dilated convolutional structure of the BiTCN captures more information through its enlarged receptive field. Furthermore, the BiLSTM enhances the retention of long-term information, and the attention mechanism dynamically emphasizes key features by weighting the feature information output by the BiLSTM. Overall, our model outperforms other models in terms of forecasting accuracy. However, due to the inherent non-smoothness of multivariate load demand series, improving the accuracy of all forecasting models remains imperative.

[0088] As a flexible resource in power system dispatch, demand-side response not only effectively reduces peak loads, alleviates grid operating pressures, and improves equipment utilization, but also promotes the adoption of new energy sources and contributes to energy conservation and emission reduction. Accurately assessing demand response potential helps identify the contribution and value of response resources to the balance of power supply and demand, thereby better assisting in optimizing grid dispatch. This has important practical significance in building a new power system and stimulating the enthusiasm of power users to participate in grid interaction.

[0089] To construct a demand-side load adjustment potential assessment model, this paper first uses the typical electricity load curve obtained through cluster analysis and prediction of load indicators in Sections 2 and 3. Furthermore, it combines the demand response load increase and decrease rates of the load industry and the actual peak-to-valley difference of the user's load to assess the theoretical potential of the user to participate in demand response, including peak shaving potential and valley filling potential. The specific steps are as follows:

[0090] Six load characteristic indicators were constructed: daily average load (A_1), daily load rate (A_2), daily peak-to-valley difference (A_3), peak load rate (A_4), flat load rate (A_5), and valley load rate (A_6). The loads were clustered using an improved fuzzy C-means algorithm. At the same time, atypical daily load data was removed to perform data dimensionality reduction.

[0091] (2) Based on the load demand forecasting model constructed in this paper, Class load forecast, get the predicted typical daily load curve of this class load .

[0092] (3) No. Type of load during peak or valley of power system The load values ​​corresponding to the time are , then the user's potential demand response peak shaving or valley filling is: #timg# （32）

[0093] Where, Indicates the The possible regulatory potential of this type of load participating in demand response, Indicates the number of hours the system peak or trough lasts. Indicates the The proportion of flexible load of this type of load under demand response measures, that is, the ratio of average load adjustment to maximum load.

[0094] (4) System peak time The difference between the actual load of the load type and the daily minimum load is The maximum value of the peak shaving potential of load type participating in demand response is The theoretical peak reduction potential of this type of load should not exceed this value; similarly, the peak reduction potential of the system should not exceed this value. The difference between the actual load of the load type and the maximum load of the day is The maximum value of the potential of this type of load to participate in demand response valley filling, The theoretical valley-filling potential of this type of load should not exceed this value. The typical load curve of the load type can be calculated The difference between the actual load of a type of load during the peak period and the daily minimum load, as well as the difference between the actual load during the valley period and the daily maximum load, are shown in the following formula: #timg# （33） #timg# （34）

[0095] Where, Indicates the peak or valley moment of the system, and its value is , , Respectively The maximum load value and the minimum load value in the load curve obtained by class load prediction.

[0096] (5) No. The potential for peak shaving and valley filling of demand response for these load types is: #timg# （35） #timg# （36）

[0097] The above content is only for explaining the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

[0098] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

[0099] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of a fully hardware embodiment, a fully software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present application may be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0100] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0101] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0102] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0103] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0104] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A load demand response characteristics analysis and prediction method based on load-side data mining, characterized by: include: 1) An improved fuzzy C-means clustering method is proposed. It is applicable to large amounts of heterogeneous demand-side power data from multiple sources. It performs dimensionality reduction preprocessing on user-side power data. This method can efficiently extract data features while reducing complexity, simplifying the construction of adjustable resource pools with construction loads, and facilitating the analysis of user consumption behavior by power grid companies. 2) A multi-type load forecasting model based on convolutional neural networks is proposed. A CNN is used to select highly relevant data and filter the set of input variables for the forecasting model. A Transformer-based load forecasting model is established and applied to load forecasting of demand-side adjustable resource pools. 3) Detailed analysis of the power consumption characteristics of various load types and the various factors influencing participation in demand response. Combined with load forecast datasets, this study identifies and quantifies the responsiveness and characteristics of demand-side adjustable resources. This study explores the multidimensional characteristics of load-side adjustable resources and unleashes their potential in regulating supply and demand. The study also examines the interactive operating characteristics and power consumption of demand-side resources.

2. The improved fuzzy C-means clustering method according to claim 1 is used to efficiently extract data features, characterized in that: Traditional clustering algorithms such as C-means and fuzzy C-means, which are commonly used for power load data preprocessing, have the problem of local convergence caused by improper selection of initial cluster centers, which affects the effect of power load preprocessing.

3. The improved fuzzy C-means clustering method according to claim 2 is used to efficiently extract data features, characterized in that: A feature preprocessing method for power load sample data based on an improved fuzzy C-means clustering algorithm is proposed. By improving the selection of cluster centers in the traditional fuzzy C-means algorithm, the power load type recognition effect is improved, laying the foundation for subsequent power load forecasting.

4. The improved fuzzy C-means clustering method according to claim 3 is used to efficiently extract data features, characterized in that: In order to improve the clustering performance of the traditional fuzzy C-means clustering algorithm for large-scale, multivariate and heterogeneous data on the load side, an improved fuzzy C-means clustering algorithm is proposed with the assistance of neural network to extract the mapping relationship between external factors such as date and weather and node daily load, and it is used for power load classification and identification.

5. The improved fuzzy C-means clustering method according to claim 4 is used to efficiently extract data features, characterized in that: Taking into account the significant differences between different load types, an improved fuzzy C-means clustering algorithm based on sparrow optimization is proposed. This algorithm classifies load samples according to climate and date characteristics, and constructs a discriminant interval between the Euclidean distance between load characteristic indicators and the improved fuzzy C-means. Data dimensionality reduction is achieved by eliminating atypical daily load curves with unclear classification characteristics.

6. Based on the improved fuzzy C-means clustering method described in claim 5, a CNN-Transformer-based power load forecasting model is designed to efficiently extract data features, which is used to verify the effectiveness of power load clustering and is also used for load forecasting of demand-side adjustable resources.

7. The CNN model according to claim 6 has two characteristics: the first characteristic is that each layer in the CNN has a local connection with the next layer; the second characteristic is that the weight values ​​are shared in the entire CNN model.

8. The CNN-Transformer-based power load forecasting model according to claim 7, characterized in that: Based on the time series feature extraction of raw data by CNN, a short-term power load forecasting method based on Transformer is further proposed. It mainly performs time series forecasting in the form of encoder-decoder, has good parallelism, and can effectively obtain the optimal results.

9. The CNN-Transformer-based power load forecasting model according to claim 8, characterized in that: The typical electricity load curve obtained by cluster analysis and prediction of load indicators is further combined with the demand response load increase and decrease rate of the industry where the load is located and the actual peak-to-valley difference of the user's load to evaluate the theoretical potential of users to participate in demand response.

Images