A multi-lstm power load short-term prediction method based on multi-scene recognition

Abstract

The application discloses a multi-LSTM power load short-term prediction method based on multi-scene recognition, and comprises the following steps: S1, collecting original data; S2, preprocessing the original data, and extracting a load scene set through a clustering algorithm; S3, constructing a load prediction model; S4, constructing a scene label prediction model; and S5, obtaining a final prediction value.The application has the beneficial effects that a plurality of differentiated scenes are mined out by collecting historical power load data from equipment, a plurality of prediction models are constructed and independently trained, the future time scene is predicted by using a scene label, a corresponding load prediction model is selected, and more accurate power load prediction is realized.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting technology, and in particular to a multi-LSTM power load short-term forecasting method based on multi-scenario recognition. Background Technology

[0002] Power system load forecasting refers to determining the required power or electricity consumption of the power system for a future period, based on past and current load data, while meeting certain accuracy requirements and comprehensively considering the operating characteristics of the power system, environmental, meteorological, and social factors. The accuracy of power load forecasting is a crucial indicator of the economic operation of the power system, directly impacting factors related to electricity costs, such as generator start-up and shutdown, maintenance arrangements, and spinning capacity reserves. Improving the accuracy of load forecasting, especially short-term load forecasting, is a key focus of current research.

[0003] Load forecasting methods generally employ statistical approaches, including the Autoregressive Integrated Moving Average (ARIMA) model, multiple regression, and exponential smoothing. These methods have relatively simple structures and struggle to uncover the complex correlations within load sequences. For example, ARIMA primarily learns the correlations within load sequences, but electricity load is typically influenced by many external factors.

[0004] Chinese patent application CN103123665A discloses a short-term power load forecasting method based on fuzzy clustering similar days. This method uses fuzzy clustering, which applies the ideas of fuzzy mathematics to cluster analysis. It is mainly applicable to the classification of qualitative variables. In this method, similar days are selected mainly based on the clustering level, but it is difficult to select appropriate similarity parameters. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a multi-LSTM power load short-term prediction method based on multi-scenario recognition.

[0006] The objective of this invention is achieved through the following technical solution: a multi-LSTM power load short-term forecasting method based on multi-scenario recognition, comprising the following steps;

[0007] S1: Collect raw data;

[0008] S2: Preprocess the raw data and extract the load scenario set using a clustering algorithm;

[0009] S3: Construct a load forecasting model;

[0010] S4: Construct a scene label prediction model;

[0011] S5: Obtain the final predicted value.

[0012] Preferably, in step S1, the raw data includes historical load data of electrical equipment, weather factors, and holiday information.

[0013] Preferably, step S2 further includes the following step:

[0014] S2.1: Sort the collected historical load data points to form complete time series data X = {x1, x2, x3, ..., x...} n Then normalize using MinMax;

[0015] S2.2: Set multiple cluster sizes, conduct repeated experiments, and calculate the silhouette coefficient of the clustering results under different numbers of clusters:

[0016]

[0017] Where, d in (x i (x) represents the i-th load data point. i The average distance to all other objects of the same type, d ic (x i ) is x i The minimum average distance to all objects of other classes, S m ∈[-1,1], when S m The larger the value, the smaller the intra-class distance, the larger the inter-class distance, and the better the clustering effect;

[0018] S2.3: Using the unsupervised clustering K-Means algorithm, various load scenarios are extracted from historical load data, and samples are clustered into K specified clusters based on the Euclidean distance of sample features.

[0019]

[0020] Where, dis(x) i C j Let x be the i-th (1≤i≤n) object. i To the j-th cluster center C j Euclidean distance (1≤j≤k);

[0021] S2.4: Select the optimal number of clusters obtained in step S2.2, and define data segments with the same label as the same scenario according to the chronological order of the clustered load data, forming a load scenario set {S1, S2, S3, ..., S...}. K}

[0022] Preferably, step S3 further includes the following step:

[0023] S3.1: Extract the temporal features of the load curve using a long short-term memory neural network. The parameter iterative update formula is as follows:

[0024] f t =σ(ω) f [a (t-1) ,x (t) ]+b f );

[0025] i t =σ(ω) i [a (t-1 ),x (t) ]+b i );

[0026] o t =σ(ω) o [a (t-1) ,x (t) ]+b o );

[0027]

[0028]

[0029] a (t) =o t *tanhC (t) M;

[0030] Among them, (f t ) is the forget gate, (i t ) is the input gate, (o t ) is the output gate;

[0031] S3.2: Select the load scenario set {S1, S2, S3, ..., S...} formed in step S2.4. K} Construct K independent LSTM prediction networks, each corresponding to a scene set, to obtain the training set.

[0032] Preferably, step S4 further includes the following step:

[0033] S4.1: Combine the climate information, holiday information and equipment load collected in step S1 to construct a scene feature dataset;

[0034] S4.2: Extract latent features of the scene using a multilayer perceptron, train the model using backpropagation, and iteratively update the network parameters;

[0035] S4.3: The scene features extracted from multiple hidden layers are input into the output layer with the Softmax activation function. A multi-classification model is trained using scene labels obtained through clustering. For n samples input into the classification training set, the corresponding labels are... The estimated probability for each sample is:

[0036]

[0037] Among them, W1, W2, W3,, W k Here are the network model parameters, and K is the number of classes. Normalize the results.

[0038] Preferably, in step S5, based on the scene labels obtained in step S4, a corresponding model is selected from the multiple prediction models trained in step S3, and historical load data is input into the prediction model to obtain the final prediction value.

[0039] The present invention has the following advantages: by collecting historical power load data from equipment, the present invention mines multiple differentiated scenarios, constructs and independently trains multiple prediction models, uses scenario labels to predict scenarios at future times, selects the corresponding load prediction model, and achieves more accurate power load prediction. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the prediction method process. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0042] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0043] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.

[0044] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0045] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use, or the orientation or positional relationship commonly understood by those skilled in the art. They are only used for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. In addition, the terms "first," "second," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0046] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0047] In this embodiment, as Figure 1 As shown, a multi-LSTM power load short-term forecasting method based on multi-scenario recognition includes the following steps;

[0048] S1: Collect raw data; further, in step S1, the raw data includes historical load data of electrical equipment, weather factors, and holiday information.

[0049] S2: Preprocess the raw data and extract the load scenario set using a clustering algorithm;

[0050] S3: Construct a load forecasting model;

[0051] S4: Construct a scene label prediction model;

[0052] S5: Obtain the final predicted value. By collecting historical power load data from equipment, multiple differentiated scenarios are mined, multiple prediction models are built and independently trained, scenario labels are used to predict scenarios at future times, and the appropriate load prediction model is selected to achieve more accurate power load prediction.

[0053] Furthermore, step S2 also includes the following steps:

[0054] S2.1: Sort the collected historical load data points to form complete time series data X = {x1, x2, x3, ..., x...} n Then normalize using MinMax;

[0055] S2.2: Set multiple cluster sizes, conduct repeated experiments, and calculate the silhouette coefficient of the clustering results under different numbers of clusters:

[0056]

[0057]

[0058] Where, d in (x i (x) represents the i-th load data point. i The average distance to all other objects of the same type, d ic (x i ) is x i The minimum average distance to all objects of other classes, S m ∈[-1,1], when S m A larger value of S indicates smaller intra-class distances, larger inter-class distances, and better clustering results; in other words, the optimal number of clusters is selected from S. m The K value corresponding to the maximum value.

[0059] S2.3: Using the unsupervised clustering K-Means algorithm, various load scenarios are extracted from historical load data, and samples are clustered into K specified clusters based on the Euclidean distance of sample features.

[0060]

[0061] Where, dis(x) i C j Let x be the i-th (1≤i≤n) object. i To the j-th cluster center C j Euclidean distance (1≤j≤k);

[0062] S2.4: Select the optimal number of clusters obtained in step S2.2, and define data segments with the same label as the same scenario according to the chronological order of the clustered load data, forming a load scenario set {S1, S2, S3, ..., S...}. K}

[0063] In this embodiment, step S3 further includes the following step:

[0064] S3.1: Extract the temporal features of the load curve using a long short-term memory neural network. The parameter iterative update formula is as follows:

[0065] f t =σ(ω) f [a( t-1) ,x (t) ]+b f );

[0066] i t =σ(ω) i [a (t-1) ,x (t) ]+b i );

[0067] o t =σ(ω) o [a (t-1) ,x (t) ]+b o );

[0068]

[0069]

[0070] a (t) =o t *tanhC (t) M;

[0071] Among them, (f t ) is the forget gate, (i t ) is the input gate, (o t The input gate is 1.5; specifically, the temporal features of the load curve are extracted through a Long Short-Term Memory (LSTM) neural network. Compared to a recurrent neural network, LSTM adds an input gate (i = 1 / 2). t ), Forget Gate (f t ) and output gate (o t The LSTM neural network outputs a hidden state at each time step. The input gate determines whether the hidden layer information is updated, the forget gate determines whether the updated information includes information from the previous time step, and the output gate determines which part of the updated information is output. Through parameter iterative update formulas, the model's output hidden state contains information from the input gate, forget gate, and output gate. Therefore, this information is extracted as the hidden feature of the scene load. Simultaneously, the LSTM neural network outputs a hidden state at each time step. The last output state is selected as the most comprehensive hidden information, and then further processed using fully connected layers to obtain the predicted value l for the i-th type of scene at a future time point. i .

[0072] S3.2: Select the load scenario set {S1, S2, S3, ..., S...} formed in step S2.4. K} Construct K independent LSTM prediction networks, each corresponding to a scene set, to obtain the training set. Specifically, taking a certain scene as an example, the corresponding prediction network slides forward with a time window of 16 sampling points to predict the load value at the 17th point. That is, input1 = {load1, load2, ..., load...} 16}, output1 = {load 17 Then, the 17th point is substituted into the input to form a new 16-dimensional input matrix to predict the 18th point. input2 = {load2, load3, ..., load...} 17}, output2 = {load 18 Therefore, the training set is composed of Input = {input1, input2, ..., input...}. (n-1)×16}, Output={output1,output2,...,output (n-1)×16} and so on, independently training the remaining prediction models, thereby effectively reducing data interference between scenarios.

[0073] Furthermore, step S4 also includes the following steps:

[0074] S4.1: Combine the climate information, holiday information and equipment load collected in step S1 to construct a scene feature dataset;

[0075] S4.2: Extract latent features of the scene using a multilayer perceptron, train the model using backpropagation, and iteratively update the network parameters;

[0076] S4.3: The scene features extracted from multiple hidden layers are input into the output layer with the Softmax activation function. A multi-classification model is trained using scene labels obtained through clustering. Specifically, the result obtained by Softmax represents the probability that the input scene temporal features are assigned to each class. Taking a K-classifier as an example, the output is a K-dimensional vector where the sum of the elements is 1. For n samples input into the classification training set, the corresponding labels are... The estimated probability for each sample is:

[0077]

[0078] Among them, W1, W2, W3,, W k Here are the network model parameters, and K is the number of classes. The results are normalized. Specifically, The results are normalized so that the sum of all probabilities is 1. In order to complete network training, gradient descent is applied to the Softmax loss function to minimize the result.

[0079] Furthermore, in step S5, based on the scene labels obtained in step S4, the corresponding model is selected from the multiple prediction models trained in step S3, and the historical load data is input into the prediction model to obtain the final prediction value.

[0080] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-LSTM power load short-term forecasting method based on multi-scenario recognition, characterized in that: Includes the following steps; S1: Collect raw data; S2: Preprocess the raw data and extract the load scenario set using a clustering algorithm; S3: Construct a load forecasting model; S4: Construct a scene label prediction model; S5: Obtain the final predicted value; In step S1, the raw data includes historical load data of electrical equipment, climate information, and holiday information; Step S2 further includes the following steps: S2.1: Sort the collected historical load data points to form complete time series data. Then normalize using MinMax; S2.2: Set multiple cluster sizes, conduct repeated experiments, and calculate the silhouette coefficient of the clustering results under different numbers of clusters: ； ； in, For the first Load data points The average distance to all other objects of the same type for The minimum average distance to all objects of other classes. ,when The larger the value, the smaller the intra-class distance, the larger the inter-class distance, and the better the clustering effect; S2.3: Using the unsupervised clustering K-Means algorithm, various load scenarios are extracted from historical load data, and samples are clustered into specified clusters based on the Euclidean distance of sample features. In each of the clusters, ； in, For the first objects To the Cluster centers The Euclidean distance, where, , ; S2.4: Select the optimal number of clusters obtained in step S2.2, and define data segments with the same label as the same scenario according to the chronological order of the clustered load data, forming a load scenario set. ; Step S3 further includes the following steps: S3.1: Extract the temporal features of the load curve using a long short-term memory neural network; S3.2: Select the set of load scenarios formed in step S2.

4. , build A training set is obtained by using 1 independent LSTM prediction network, with each prediction network corresponding to a scene set; Step S4 also includes the following steps: S4.1: Combine the climate information, holiday information and historical load data collected in step S1 to construct a scene feature dataset; S4.2: Extract latent features of the scene using a multilayer perceptron, train the model using backpropagation, and iteratively update the network parameters; S4.3: The scene features extracted from multiple hidden layers are input into the output layer with the Softmax activation function. A multi-classification model is trained using the scene labels obtained through clustering. When a sample is input into the classification training set, its corresponding label is... The estimated probability for each sample is: ； in, , , , , These are all network model parameters. For the number of categories, Normalize the results.

2. The multi-LSTM power load short-term forecasting method based on multi-scenario recognition according to claim 1, characterized in that: In step S5, based on the scene labels obtained in step S4, a corresponding model is selected from the multiple prediction models trained in step S3, and historical load data is input into the prediction model to obtain the final prediction value.

Images