Short-term load prediction method considering load multi-component characteristics
By employing STL decomposition and multi-model forecasting methods, the problem of low accuracy in short-term load forecasting is solved, achieving high-precision forecasting of load sequences and ensuring the stability and efficiency of the power system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-25
- Publication Date
- 2026-04-07
AI Technical Summary
Short-term load forecasting accuracy is low because existing methods fail to fully consider the trends, periodicity, and volatility of load sequences, resulting in insufficient forecast accuracy.
The load data is decomposed into trend components, periodic components, and residual components using STL decomposition. LSTM, CNN, and GPR models are then used for prediction, and the prediction results are superimposed to obtain the final load prediction result.
It improves the accuracy of short-term load forecasting, provides an accurate understanding of future electricity demand and patterns, and maintains the stable operation of the power system.
Smart Images
Figure CN121813299A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to short-term load forecasting in the power industry, in particular to a short-term load forecasting method considering the multi-component characteristics of load. BACKGROUND
[0002] The main purpose of the power system is to meet the power demand. Load forecasting provides a clear understanding of future demand and patterns, which is the basis for planning and decision-making process. By predicting future power demand, power companies can reasonably plan power generation, transmission and distribution facilities, so as to avoid power shortage and reduce power outage. Such prediction also helps to optimize resource allocation, reduce operating costs, and improve overall system efficiency. In short, load forecasting is an indispensable part of power system planning and scheduling, which is related to the stability and sustainability of power supply, and has important influence on economic and social development.
[0003] Therefore, load forecasting has always been a hot issue in the field of power. Load forecasting is divided into traditional classical forecasting methods and intelligent forecasting methods. The traditional classical forecasting method based on statistical theory is widely used, which can include time series analysis, regression analysis, etc. These methods mainly rely on historical data and mathematical models for prediction. With the development of intelligent algorithms, power load forecasting methods based on some intelligent algorithms have appeared. These algorithms include support vector machines, long short-term memory neural networks, gated recurrent unit networks and convolutional neural networks, which can better handle nonlinear relationships and complex features.
[0004] However, the high complexity of load sequence makes the precision of load forecasting limited by a single algorithm. In order to improve the accuracy of load forecasting, signal decomposition technology can be applied in the field of load forecasting. Signal decomposition technology decomposes the load sequence into multiple components, and each component is predicted separately to improve the feature extraction and prediction accuracy of the load sequence. SUMMARY
[0005] The present application mainly aims at the problem of difficult short-term load forecasting and low precision, and proposes a short-term load forecasting method considering the multi-component characteristics of load. The method uses STL decomposition to decompose the original load data into trend component, periodic component and residual component. According to the characteristics of each component, LSTM model, CNN model and GPR model are established respectively to predict the components respectively, and then the final load forecasting result is obtained by superposition. The method is used for short-term load forecasting, which provides understanding of future power demand and patterns, and maintains stable operation of power system.
[0006] The purpose of the present application is achieved by the following technical scheme: a short-term load forecasting method considering the multi-component characteristics of load, comprising the following steps:
[0007] Comprising the following steps:
[0008] 1) Collecting the time series data of the historical load condition of the local power system, and decomposing the collected time series data into periodic component, trend component and residual component through STL decomposition; part of the periodic component, trend component and residual component obtained by decomposition is taken as a training set, and the other part is taken as a prediction set;
[0009] 2) Constructing a long short-term memory network LSTM model, a convolutional neural network CNN model and a Gaussian process regression GPR model for load prediction;
[0010] 3) Training the LSTM model with the periodic component in the training set, training the CNN model with the trend component in the training set, and training the GPR model with the residual component in the training set, to obtain the LSTM model, the CNN model and the GPR model for load prediction;
[0011] 4) Inputting the periodic component, trend component and residual component in the prediction set into the trained LSTM model, CNN model and GPR model; using the trained LSTM model to predict the load of the trend component to obtain the trend component prediction result; using the trained CNN model to predict the load of the periodic component to obtain the periodic component prediction result; using the trained GPR model to predict the load of the residual component to obtain the residual component prediction result;
[0012] 5) Superimposing the trend component prediction result, the periodic component prediction result and the residual component prediction result obtained in step 4) to obtain the final load prediction result.
[0013] As a preferred scheme of the present application, in step 1), the collected time series data is decomposed into three components of periodic component, trend component and residual component through STL decomposition, the periodic component includes the periodic change of the time series data; the trend component captures the long-term trend of the time series data; and the residual component contains random fluctuations or noise in the time series data; the relationship of each component and the time series data is as follows.
[0014] Y t =S t +T t +R t
[0015] Wherein: Y t is the value of the time series data at time t; S t is the value of the periodic component at time t; T t is the value of the trend component at time t; R t is the value of the residual component at time t.
[0016] As a preferred scheme of the present application, in step 1), when dividing the training set and the prediction set, the part of the periodic component, the trend component and the residual component in time sequence in front is taken as the training set, and the other part is taken as the prediction set.
[0017] As a preferred scheme of the present application, the STL decomposition uses the local weighted regression Loess to extract the smooth estimation of the three components: the STL decomposition is based on Loess, which consists of two recursive processes of inner loop and outer loop, the inner loop is based on Loess, and the smooth processing is performed on the periodic component and the trend component; the outer loop is based on the result of the residual component, and the robust weight is calculated to reduce the influence of the abnormal value of the time sequence data on the residual value; the calculation formula of the robust weight is as follows:
[0018]
[0019] Wherein: p t is the robust weight at time t; median(·) is the median function; B(a) is the Bisquare function; a is the independent variable of the Bisquare function; R t is the value of the residual component at time t.
[0020] In the next inner loop, the neighborhood weight in Loess is updated based on the robust weight, and the updating mode is as follows:
[0021] υ t,k+1 = υ t,k p t
[0022] Wherein: υ t,k is the neighborhood weight at time t in the kth inner loop.
[0023] The present application also provides a short-term load forecasting system for realizing the short-term load forecasting method, comprising:
[0024] A time sequence data acquisition module is configured to acquire time sequence data of historical load conditions of a local power system.
[0025] A time sequence data decomposition module is configured to decompose the acquired time sequence data into a periodic component, a trend component and a residual component through STL decomposition.
[0026] A model construction and training module is configured to construct a long short-term memory network LSTM model, a convolutional neural network CNN model and a Gaussian process regression GPR model, and train the CNN model by taking the decomposed trend component as a training set, and train the GPR model by taking the decomposed residual component as a training set, to obtain the LSTM model, the CNN model and the GPR model which can be used for load forecasting.
[0027] The load forecasting result acquisition module is used to perform load forecasting on the trend component using a trained LSTM model to obtain the trend component forecasting result; to perform load forecasting on the periodic component using a trained CNN model to obtain the periodic component forecasting result; to perform load forecasting on the residual component using a trained GPR model to obtain the residual component forecasting result; and to superimpose the trend component forecasting result, the periodic component forecasting result, and the residual component forecasting result to obtain the final load forecasting result.
[0028] Compared with the prior art, the present invention has the following beneficial effects:
[0029] This invention provides accurate short-term load forecasting. The high complexity of load sequences limits the accuracy of single-algorithm load forecasting. While signal decomposition techniques have been widely used in load forecasting in recent years, they do not consider the specific characteristics of the load and focus more on the decomposition of high-frequency and low-frequency components, resulting in poor interpretability. The strategy proposed in this invention utilizes STL to decompose the original load sequence, fully considering the load's trend, periodicity, and volatility.
[0030] This invention addresses the problem of poor prediction accuracy caused by existing methods using the same model for all components, which ignores the different characteristics of trend components, periodic components, and residual components. It proposes to use a model that matches the characteristics of each component for individual prediction, thereby improving the feature extraction and prediction accuracy of load sequences. Attached Figure Description
[0031] Figure 1 This is a flowchart of the prediction model of the present invention.
[0032] Figure 2 This is a graph showing the STL decomposition results of the first two weeks of the load sequence in this invention.
[0033] Figure 3 This is a graph showing the 50-hour load prediction results of this invention. Detailed Implementation
[0034] The present invention will be further described and illustrated below with reference to specific embodiments. The embodiments described are merely examples of the content of this disclosure and do not limit the scope of the invention. The technical features of each embodiment in the present invention can be combined accordingly, provided that there is no mutual conflict.
[0035] This invention mainly addresses the problems of high difficulty and low accuracy in short-term load forecasting. It proposes a method that considers multiple loads: 1) Collecting time-series data of historical load conditions of the local power system and decomposing the collected time-series data into periodic components, trend components, and residual components through STL decomposition; part of the decomposed periodic components, trend components, and residual components is used as the training set, and the other part is used as the prediction set.
[0036] 2) Construct a Long Short-Term Memory (LSTM) network model, a Convolutional Neural Network (CNN) model, and a Gaussian Process Regression (GPR) model for load forecasting;
[0037] 3) Train the LSTM model using the periodic components from the training set, train the CNN model using the trend components from the training set, and train the GPR model using the residual components from the training set to obtain the LSTM model, CNN model, and GPR model for load prediction.
[0038] 4) Input the periodic component, trend component, and residual component from the prediction set into the trained LSTM model, CNN model, and GPR model; use the trained LSTM model to perform load prediction on the trend component to obtain the trend component prediction result; use the trained CNN model to perform load prediction on the periodic component to obtain the periodic component prediction result; use the trained GPR model to perform load prediction on the residual component to obtain the residual component prediction result.
[0039] 5) The trend component prediction results, periodic component prediction results and residual component prediction results obtained in step 4) are superimposed to obtain the final load prediction results.
[0040] In one specific embodiment of the present invention, measured power load data from a province in northern China from January 1, 2021 to December 31, 2021 was used as the experimental dataset, with a sampling interval of 1 hour, totaling 8760 samples. The training set and test set were divided in a 4:1 ratio, with 7008 load data samples from the early stage serving as the training set and 1752 load samples from the later stage serving as the prediction set.
[0041] We constructed a Long Short-Term Memory (LSTM) network model, a Convolutional Neural Network (CNN) model, and a Gaussian Process Regression (GPR) model for load forecasting.
[0042] The Long Short-Term Memory (LSTM) network model includes memory cells, an input gate, a forget gate, and an output gate. The memory cells store input information. The input gate writes the current input information into the memory cell, and the forget gate forgets the information in the memory cell. The output gate uses the information from the memory cell as the current output. The specific expression of the LSTM network model is as follows:
[0043]
[0044] Among them, c t It is the cell state information at time t; h t It is the hidden state information at time t; x t It is the input at time t; σ and tanh are activation functions; f tThis is the result of calculating the state of the forget gate at time t; i t This is the result of calculating the state of the input gate at time t; It is the candidate cell state information at time t; o t This is the result of calculating the state of the output gate at time t; W f W is the weight matrix of the forget gate; i W is the weight matrix of the input gate; o W is the weight matrix of the output gate; c It is the weight matrix of the memory units; b f It is the bias of the forgetting gate; b i It is the input gate bias; b c It is the bias of the input gate tanh activation function; b o It is the bias of the output gate.
[0045] A Convolutional Neural Network (CNN) model includes an input layer responsible for inputting the original data, a convolutional layer that extracts features from the input data through convolution operations, a pooling layer responsible for reducing the data dimensionality of the data after processing by the convolutional layers, a fully connected layer that aggregates the input data features obtained after pooling operations, and an output layer responsible for outputting the final result.
[0046] The Gaussian Process Regression (GPR) model treats data points as a whole Gaussian process. The GPR model includes a mean function and a covariance function; the mean function defines the average level of the Gaussian process, while the covariance function defines the similarity between any two data points.
[0047] f(x)~GP(m(x),k(x,x′))
[0048] Where f(x) is a random function defined at the input point x; m(x) is the mean function; and k(x,x′) is the covariance function.
[0049] In the GPR model, the hyperparameters in the covariance function are automatically optimized by maximizing the log-marginal likelihood function; the covariance function is chosen as a Gaussian kernel function, as shown in the following equation:
[0050]
[0051] Where k(x,x′) represents the kernel function, used to calculate the similarity between two sample points x and x′; σ 2 is the variance parameter, which determines the overall range of function value variation; l is the length scale parameter, which controls the rate at which the function value changes with the distance between input points.
[0052] STL was used to decompose 8760 load data samples. STL decomposition uses locally weighted regression (Loess) to extract smoothed estimates of the three components. Based on Loess, STL decomposition consists of two recursive processes: an inner loop and an outer loop. The inner loop, based on Loess, smooths the periodic and trend components. The outer loop calculates robust weights based on the residual component results to reduce the impact of outliers on the residuals. The formula for calculating the robust weights is shown below:
[0053]
[0054] Where: ρ t Let be the robust weight at time t; median(·) is the median function; B(a) is the Bisquare function; a is the independent variable of the Bisquare function; R t It is the value of the residual component at time t.
[0055] In the next inner loop, the neighborhood weights in Loess will be updated based on the robust weights, as shown in the following formula:
[0056] υ t,k+1 =υ t,k ρ t
[0057] Where: υ t,k Let be the neighborhood weight at time t during the k-th inner loop.
[0058] The periodic components, trend components, and residual components obtained from the decomposition are used, with one portion serving as the training set and the other as the test set. The model is trained using the training set, and component predictions are performed based on the trained model. Specifically:
[0059] For relatively gentle trend components, an LSTM model is used. In the LSTM model, the time step is set to 24, meaning the value of the 25th trend component is predicted using the values of the previous 24 historical trend components. The LSTM has 3 layers, each with 64 units. The activation function is tanh, the loss function is mse, the maximum number of iterations is set to 1000, and the loss convergence is set to 0.000001. The LSTM model is then trained using a training set: the weights and biases in the LSTM model are randomly initialized; the training set data is input into the LSTM model, the model's predicted output for each input is calculated, and the mean squared error function is used to calculate the difference between the model's predicted value and the true value; the gradient of the loss function with respect to the weights and biases in the model parameters is calculated using the backpropagation algorithm; the weights and biases in the model are updated using the gradient descent algorithm to minimize the loss function; these steps are repeated until the set stopping condition is met, i.e., the maximum number of iterations or loss convergence is achieved. The prediction set samples are then input into the trained model for prediction to obtain the prediction results.
[0060] For the periodic components, a CNN model is used. In the CNN model, both convolutional layers use 32 filters with a filter size of 3 and a stride of 1, using 'same' padding, and the activation function is tanh. A max-pooling layer with a pooling size of 2 is added between the two convolutional layers. The fully connected layer has 50 neurons with the activation function tanh. The output layer has 1 neuron. The loss function is MSE, the maximum number of iterations is set to 1000, and the loss convergence parameter is set to 0.000001. Next, the CNN model is trained using the training set: the convolutional kernels, biases, and weights in the CNN model are randomly initialized; the training set data is input into the CNN model, and after convolution and pooling operations, the predicted output is obtained, and the mean squared error function is used to calculate the difference between the model's predicted value and the true value; the gradient of the loss function with respect to the convolutional kernels, biases, and weights is calculated using the backpropagation algorithm; according to the gradient descent algorithm, the convolutional kernels, biases, and weights in the model are updated to minimize the loss function; the above steps are repeated until the set stopping condition is met, i.e., the maximum number of iterations or loss convergence is reached. The prediction set samples are then input into the trained model for prediction to obtain the prediction results.
[0061] For residual components with large fluctuations, the GPR model is used. In the GPR model, the length scale is set to 1, the noise level is set to 0.3, the mean function is set to 0, the maximum number of iterations is set to 1000, and the loss convergence is set to 0.000001.
[0062] Next, the GPR model is trained using the training set: the variance of the Gaussian kernel function is initialized; the gradient of the variance is obtained by differentiating the marginal likelihood function; the search direction of descent is determined using the conjugate gradient method, and the variance of the kernel function is updated according to the search direction; the above steps are repeated until the set stopping condition is met, that is, the maximum number of iterations is reached or the change in variance is less than the loss convergence. The prediction set samples are then input into the trained model for prediction to obtain the prediction results.
[0063] The results of the component predictions are superimposed to obtain the final load prediction result: Y t =S t +T t +R t ,
[0064] Where: Y t It is the value of the time series data at time t; S t It is the value of the periodic component at time t; T t It is the value of the trend component at time t; R t It is the value of the residual component at time t.
[0065] like Figure 3 The image shows a comparison between the load forecasting results obtained using this invention and the actual values. The method of this invention can accurately predict short-term loads, providing an understanding of future electricity demand and patterns, and maintaining the stable operation of the power system.
[0066] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. Those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.
Claims
1. A short-term load forecasting method considering the multi-component characteristics of load, characterized in that, Includes the following steps: 1) Collect time-series data of historical load conditions of the local power system, and decompose the collected time-series data into periodic components, trend components and residual components through STL decomposition. The periodic components, trend components, and residual components obtained from the decomposition are divided into a training set and a prediction set, respectively. 2) Construct a Long Short-Term Memory (LSTM) network model, a Convolutional Neural Network (CNN) model, and a Gaussian Process Regression (GPR) model for load forecasting; 3) Train the LSTM model using the periodic components from the training set, train the CNN model using the trend components from the training set, and train the GPR model using the residual components from the training set to obtain the LSTM model, CNN model, and GPR model for load prediction. 4) Input the periodic component, trend component, and residual component from the prediction set into the trained LSTM model, CNN model, and GPR model; use the trained LSTM model to perform load prediction on the trend component to obtain the trend component prediction result; use the trained CNN model to perform load prediction on the periodic component to obtain the periodic component prediction result; use the trained GPR model to perform load prediction on the residual component to obtain the residual component prediction result. 5) The trend component prediction results, periodic component prediction results and residual component prediction results obtained in step 4) are superimposed to obtain the final load prediction results.
2. The short-term load forecasting method according to claim 1, characterized in that, In step 1), the collected time series data is decomposed into three components—periodic component, trend component, and residual component—through STL decomposition. The periodic component includes the periodic changes of the time series data. The trend component captures the long-term trend of time series data; while the residual component contains random fluctuations or noise in the time series data; the relationship between each component and the time series data is as follows: Y t =S t +T t +R t Where: Y t It is the value of the time series data at time t; S t It is the value of the periodic component at time t; T t It is the value of the trend component at time t; R t It is the value of the residual component at time t.
3. The short-term load forecasting method according to claim 2, characterized in that, STL decomposition uses locally weighted regression Loess to extract smooth estimates of the three components: STL decomposition is based on Loess and consists of two recursive processes, an inner loop and an outer loop. The inner loop is based on Loess and smooths the periodic and trend components. The outer loop calculates robust weights based on the results of the residual components to reduce the impact of outliers in the time series data on the residuals; the formula for calculating the robust weights is: Where: ρ t Let be the robust weight at time t; median(·) is the median function; B(a) is the Bisquare function; a is the independent variable of the Bisquare function; R t It is the value of the residual component at time t; In the next inner loop, the neighborhood weights in Loess will be updated based on the robust weights, as shown in the following formula: u t,k+1 =υ t,k r t Where: υ t,k Let υ be the neighborhood weight at time t during the k-th inner loop. t,k+1 Let be the neighborhood weight at time t during the (k+1)th inner loop.
4. The short-term load forecasting method according to claim 1, characterized in that, In step 2), the Long Short-Term Memory (LSTM) network model includes memory cells, an input gate, a forget gate, and an output gate. The memory cells are used to store input information. The input gate is used to write the current input information into the memory cells, and the forget gate is used to forget the information in the memory cells. The output gate is used to use the information from the memory cells as the current output. The specific expression of the LSTM network model is as follows: Among them, c t It is the cell state information at time t; h t It is the hidden state information at time t; x t It is the input at time t; σ and tanh are activation functions; f t This is the result of calculating the state of the forget gate at time t; i t This is the result of calculating the state of the input gate at time t; It is the candidate cell state information at time t; t This is the result of calculating the state of the output gate at time t; W f W is the weight matrix of the forget gate; i W is the weight matrix of the input gate; o W is the weight matrix of the output gate; c It is the weight matrix of the memory units; b f It is the bias of the forgetting gate; b i It is the bias of the input gate; b c It is the bias of the input gate tanh activation function; b o It is the bias of the output gate.
5. The short-term load forecasting method according to claim 1, characterized in that, In step 2), the Convolutional Neural Network (CNN) model includes an input layer responsible for inputting the original data, a convolutional layer that extracts features from the input data through convolution operations, a pooling layer responsible for reducing the data dimensionality of the data after processing by the convolutional layers, a fully connected layer that aggregates the input data features obtained after pooling operations, and an output layer responsible for outputting the final result.
6. The short-term load forecasting method according to claim 1, characterized in that, In step 2), the Gaussian Process Regression (GPR) model is used to treat the data points as a whole Gaussian process. The GPR model includes a mean function and a covariance function; the mean function is used to define the average level of the Gaussian process, while the covariance function is used to define the similarity between any two data points. f(x)~GP(m(x),k(x,x′)) Where f(x) is a random function defined at the input point x; m(x) is the mean function; and k(x,x′) is the covariance function.
7. The short-term load forecasting method according to claim 6, characterized in that, In the GPR model, the hyperparameters in the covariance function are automatically optimized by maximizing the log-marginal likelihood function; the covariance function is chosen as a Gaussian kernel function, as shown in the following equation: Where k(x,x′) represents the kernel function, used to calculate the similarity between two sample points x and x′; σ 2 is the variance parameter, which determines the overall range of function value variation; l is the length scale parameter, which controls the rate at which the function value changes with the distance between input points.
8. The short-term load forecasting method according to claim 6, characterized in that, In step 3), training the GPR model with the training set specifically involves: initializing the variance of the Gaussian kernel function; taking the derivative of the variance based on the marginal likelihood function to obtain its gradient; determining the descent search direction using the conjugate gradient method and updating the kernel function variance based on the search direction; and repeating the above steps until the set stopping condition is met, i.e., reaching the maximum number of iterations or the change in variance being less than the preset loss convergence.
9. The short-term load forecasting method according to claim 1, characterized in that, In step 3), training the LSTM model with the training set specifically involves: initializing the weights and biases in the LSTM model, inputting the training set data into the LSTM model, the LSTM model predicting output for each input, and using the mean squared error function to calculate the difference between the model's predicted value and the true value. The gradient of the loss function with respect to the weights and biases in the model parameters is calculated using the backpropagation algorithm; the weights and biases in the model are updated according to the gradient descent algorithm; the above steps are repeated until the maximum number of iterations is reached or the loss converges to less than or equal to a preset value; the training method of the CNN model is the same as that of the LSTM model.
10. A short-term load forecasting system for implementing the short-term load forecasting method of claim 1, characterized in that, include: The time-series data acquisition module is used to collect time-series data on the historical load of the local power system. The time series data decomposition module is used to decompose the collected time series data into periodic components, trend components, and residual components through STL decomposition. The decomposed periodic components, trend components, and residual components are divided into training sets and prediction sets. The model building and training module is used to build a Long Short-Term Memory (LSTM) network model, a Convolutional Neural Network (CNN) model, and a Gaussian Process Regression (GPR) model. The trend components obtained from the decomposition are used as the training set to train the CNN model, and the residual components obtained from the decomposition are used as the training set to train the GPR model, resulting in LSTM, CNN, and GPR models that can be used for load forecasting. The load forecasting result acquisition module is used to perform load forecasting on the trend component using a trained LSTM model to obtain the trend component forecasting result; to perform load forecasting on the periodic component using a trained CNN model to obtain the periodic component forecasting result; and to perform load forecasting on the residual component using a trained GPR model to obtain the residual component forecasting result. The trend component prediction results, periodic component prediction results, and residual component prediction results are then superimposed to obtain the final load prediction result.