Long-term power load prediction method, device and system, and storage medium
The CEEMD-LSTM model decomposes historical power load data and performs deep learning, solving the accuracy problem of long-term power load prediction and achieving higher prediction accuracy.
Patent Information
- Application Number
- CN202411327337.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-23
- Publication Date
- 2025-01-07
AI Technical Summary
The prior art is difficult to accurately predict long-term power loads, and short-term prediction methods are mature but cannot effectively predict future power loads.
Using the CEEMD-LSTM model, the future power load size is predicted by decomposing historical long-term power load data into multiple eigenmodal function components and using long and short-term memory neural networks for deep learning training.
The accuracy of power load prediction is improved, and the accuracy of the prediction model is improved by reducing data volatility and reflecting time series feature information.
Smart Images
Figure CN119272002A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of power systems, and in particular relates to a long-term power load forecasting method and device, system, and storage medium. Background Art
[0002] With the advent of the electrical age, people have gradually invested more and more energy in the power industry. With the increasing application of electricity, the use of electricity has become more and more popular, and the power system has gradually developed and improved, people have begun to independently study the load forecasting of the power system. Due to the small amount of electricity production and manufacturing in the initial stage of power resource application, the quality is not high enough, the resources are tight and even in short supply. In order to plan the distribution of power equipment in the future and formulate a reasonable power supply construction plan for power production, load forecasting has become a difficult problem that must be overcome for the company's business development. With the investment of funds, the rapid development of science and technology, and the hard work of technicians, new forecasting methods continue to emerge, and load forecasting has gradually developed. Power load forecasting is based on previous load data, through some models to summarize the rules hidden under the numbers, and then estimate the load size for a period of time in the future. In terms of load forecasting, the short-term load forecasting method system has matured, but it is impossible to predict long-term power load. Summary of the invention
[0003] The technical problem to be solved by the present invention is to provide a long-term power load forecasting method and device, system, and storage medium, which can improve the prediction accuracy.
[0004] To achieve the above object, the present invention adopts the following technical solution:
[0005] A long-term power load forecasting method, comprising:
[0006] Step S1, obtaining historical long-term power load data;
[0007] Step S2: training the CEEMD-LSTM model based on historical long-term power load data;
[0008] Step S3: predicting the power load size at a predetermined time in the future according to the trained CEEMD-LSTM model; wherein,
[0009] Step S2 includes:
[0010] The historical long-term power load data is decomposed into multiple intrinsic mode function components through CEEMD;
[0011] According to multiple intrinsic mode function components IMF, multiple long short-term memory neural network models are trained for deep learning respectively;
[0012] Step S3 includes:
[0013] Predict future scheduled time data through multiple trained long short-term memory neural network models to obtain multiple IMF component prediction values;
[0014] Reconstruct the forecast values of multiple IMF components to obtain the final forecast result.
[0015] The present invention also provides a long-term power load forecasting device, comprising:
[0016] An acquisition module is used to obtain historical long-term power load data;
[0017] The training module is used to train the CEEMD-LSTM model based on historical long-term power load data;
[0018] The prediction module is used to predict the power load size at a predetermined time in the future based on the trained CEEMD-LSTM model;
[0019] The training modules include:
[0020] A decomposition unit, used for decomposing the historical long-term power load data into multiple intrinsic mode function components through CEEMD;
[0021] A training unit, used for performing deep learning training on a plurality of long short-term memory neural network models respectively according to a plurality of intrinsic mode function components;
[0022] The prediction module includes:
[0023] A prediction unit, used to predict future scheduled time data through multiple trained long short-term memory neural network models to obtain multiple IMF component prediction values;
[0024] The reconstruction unit is used to reconstruct the prediction values of multiple IMF components to obtain the final prediction result.
[0025] The present invention also provides a long-term power load forecasting system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a long-term power load forecasting method when executed by the processor.
[0026] The present invention also provides a storage medium, on which a computer program is stored, and the computer program executes the long-term power load forecasting method when running.
[0027] The present invention uses CEEMD to deal with the problem of non-stationary data. The IMF components decomposed according to different frequencies can reduce the volatility of the data. The LSTM neural network model can reflect the characteristic information of the time series and connect the different relationships of various data in the long and short time. The technical solution of the present invention can improve the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0029] Figure 1 A flowchart of a long-term power load forecasting method according to an embodiment of the present invention;
[0030] Figure 2 This is a schematic diagram of a long-term power load forecasting method according to an embodiment of the present invention;
[0031] Figure 3 Schematic diagram of actual values and predicted values of CEEMD-LSTM model. DETAILED DESCRIPTION
[0032] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0033] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0034] Embodiment 1:
[0035] like Figure 1 , 2 As shown, an embodiment of the present invention provides a long-term power load forecasting method, comprising:
[0036] Step S1, obtaining historical long-term power load data;
[0037] Step S2: training the CEEMD-LSTM model based on historical long-term power load data;
[0038] Step S3: predict the power load size at a predetermined time in the future based on the trained CEEMD-LSTM model.
[0039] As an implementation of the embodiment of the present invention, step S2 includes:
[0040] Step S21, decomposing the historical long-term power load data into multiple intrinsic mode function components through CEEMD;
[0041] Step S22: performing deep learning training on multiple long short-term memory neural network models according to multiple intrinsic mode function components.
[0042] Further, in step S21, the CEEMD analysis method adds a pair of mutually opposite positive and negative white noises as auxiliary noises to the historical long-term power load data (i.e., the original data), and derives two mixed matrices consisting of the original data and the added noise in the following manner:
[0043]
[0044] Among them, X is the original data, N is the added white noise, M1 is the sum of the original data and positive noise, and M2 is the sum of the original data and negative noise.
[0045] A new set of mixed time series ((M1+M2) / 2) is obtained. After performing empirical mode decomposition (EMD) on each time series, the historical long-term power load data is finally decomposed into a finite number of IMF components through multiple decompositions.
[0046]
[0047] Among them, c i,j (t) is the jth IMF obtained by decomposition after adding positive and negative noise for the i-th time, r i,j (t) is the residual function.
[0048] Since the power load time series of the historical long-term power load data includes M power load measurements, at any decomposition level, each intrinsic mode function component can be regarded as a power load component obtained by decomposing the power load measurements of all sampling points of the power load time series. The number of power load components is equal to the number of sampling points of the power load time series (M). That is, the time series length of each intrinsic mode function component (i.e., the number of power load components) is equal to M, each subsequence in the intrinsic mode function is a power load component, and the sampling point corresponding to each intrinsic mode function component is consistent with the obtained power load time series, and each power load component is decomposed by the power load measurement value at the corresponding sampling point.
[0049] Further, in step S22, for any decomposition level, taking the first intrinsic mode function component IMF1 as an example, the number of training sets is equal to T, the length of the time series of the first intrinsic mode function component is equal to M, where M > T, and the number of test sets is equal to M - T. When the input of the LSTM model is the first intrinsic mode function component IMF1, the first intrinsic mode function component IMF1 includes the current sampling point and M - 1 historical sampling points. The M time series are normalized and set with a time step of Step to create a two-dimensional time series dataset as a sample (training set: from Step to T; test set: from Step to M - T). Its training set dimension is T × Step and the test set dimension is (M - T) × Step. The time step can be defined according to the designed requirements (Step < M - T < T). The created two-dimensional time series dataset is converted into a three-dimensional time series dataset (training set dimension: T × Step × 1; test set dimension: (M - T) × Step × 1). The three-dimensional time series training set is used as the input of the LSTM to obtain the trained prediction model. Then, the three-dimensional time series test set is input into the prediction model trained by the LSTM, and finally, the power load prediction value set is obtained, with a dimension of (M - T) × 1. The power load prediction value set is converted into the final power load prediction value set data after inverse normalization, and its data length is M - T. The power load component of this prediction value set will be used as the power load component prediction value set of the first intrinsic mode function component IMF1. This power load component prediction value set includes M - T power load component prediction values, and the actual values corresponding to the M - T power load component prediction values are the power load components at the current sampling point and the previous M - T - 1 historical sampling points in the first intrinsic mode function component IMF1. Figure 3 For the results of the predicted value (M - T) and the actual value (M - T) after training. To verify the feasibility of the LSTM prediction model, three indicators are proposed to verify the accuracy of the LSTM prediction model. First, the mean absolute percentage error (MAPE) is an indicator to measure the accuracy of the prediction model, especially suitable for evaluating the performance of time series prediction models. The mathematical expression is: MAPE = (1 / n) × Σ|(actual value - predicted value) / actual value| × 100%, where n represents the number of samples and Σ represents summation. The smaller the value of MAPE, the higher the accuracy of the prediction model. Second, the coefficient of determination (R - Square, R2) determines the degree of correlation. When R2 is closer to 1, it indicates that the reference value of the relevant predicted value is higher; on the contrary, when it is closer to 0, it indicates a lower reference value. The mathematical expression is: R2 = 1 - (sum of squared residuals / total variation). Third, the correlation coefficient is a statistical indicator used to reflect the degree of correlation between variables. The closer the absolute value of the correlation coefficient is to 1, the higher the degree of correlation between the actual value and the predicted value, that is, the more similar these two variables are. The mathematical expression is: Where n is the number of samples; ∑xy is the product of x and y; ∑x is the sum of x; ∑y is the sum of y; ∑x 2 represents the sum of the squares of x; ∑y 2 Represents the sum of squares of y; x represents the actual value, and y represents the predicted value. Table 1 lists the relevant quantitative indicators of the LSTM and CEEMD-LSTM models. The mean absolute percentage errors of LSTM and CEEMD-LSTM are close to zero, indicating that the predicted results are close to the actual values, but the results obtained by CEEMD-LSTM are more accurate. The R2 results are close to 1. The reference value of the predicted value of CEEMD-LSTM is relatively high compared with the predicted value of LSTM, highlighting that this method has a high degree of improvement in the prediction model. According to the results of the correlation coefficient in Table 1, the predicted value of CEEMD-LSTM is closer to the actual value than the predicted value of LSTM. According to the results of the relevant quantitative indicators in Table 1, the CEEMD-LSTM model has the advantage of improving the model accuracy.
[0050] Table 1
[0051] Quantitative indicators LSTM Model CEEMD-LSTM model Mean Absolute Percentage Error (MAPE) 2.2414％ 1.8982％ Coefficient of determination (R-Square, R2) 0.93774 0.99198 Correlation coefficient (R) 0.95753 0.96800
[0052] The core idea of LSTM is to control the flow of information through three gates (input gate, forget gate, and output gate). These gates map the input to a value between 0 and 1 through the sigmoid function (σ), thereby deciding whether to retain or discard the information.
[0053]
[0054] Among them, W i , W f , W C , W o is the weight matrix; b i 、b f 、b C 、b o is the bias term; x t is the input at the current moment, h t-1 is the hidden state of the previous moment; i t is the input gate output; f t is the forget gate output; is the information brought by the new input; C t is the cell state at the current moment; o t is the output gate output; h t is the output of the hidden state at time t and the input of the next state. The learning process of the LSTM model is a process of backpropagating errors and learning weights. The first gate is the forget gate, and the hidden state h of the previous time step is t-1 and the input information x at time t tsplicing, enter the neural network layer of the sigmoid activation function, and generate f t Data, representing the cell state C t-1 The ratio that should be retained. 0 represents C t-1 All data is forgotten, 1 represents C t-1 All data are retained. The trainable parameters are W f 、b f Parameter matrix. The second gate input gate, the data h of the previous time step t-1 and the input information x at time t t splicing, enter the neural network layer of the sigmoid activation function, and generate i t Data; the hidden state h of the previous time step t-1 and the input information x at time t t Splicing, entering the neural network layer of the tanh activation function, generates Data, will i t and The result vector is multiplied by the dot, and finally decides whether to update the cell state C t The trainable parameter is W i 、b i , W C 、b C Parameter matrix. The third gate output gate, the hidden state h of the previous time step t-1 and the input information x at time t t After concatenation and passing through a neural network layer with a sigmoid activation function, we get a vector o t , which will determine the probability of outputting data. t After a layer of tanh activation function neural network, a vector between -1 and 1 is obtained, which is the candidate output data. t The probability vector of the candidate is multiplied by the output result vector, and the final output result h is obtained. t The trainable parameter is W o 、b o Parameter matrix.
[0055] As an implementation of the embodiment of the present invention, step S3 includes:
[0056] Step S31, predicting future scheduled time data through multiple trained long short-term memory neural network models to obtain multiple IMF component prediction values;
[0057] Step S32: reconstruct the predicted values of multiple IMF components to obtain the final prediction result.
[0058] Further, in step S31, for any number of decomposition levels, the corresponding IMF component prediction values are obtained based on all intrinsic mode function components using the trained LSTM neural network model, and the IMF component prediction values are a set of power load component prediction values.
[0059] Further, in step S32, the N power load forecast value sets are averaged to obtain a new set of power load forecast value set c(t), whose length is MT.
[0060] where c j (t) is the jth IMF obtained by decomposition after adding positive and negative noise.
[0061] Embodiment 2:
[0062] The embodiment of the present invention further provides a long-term power load forecasting device, comprising:
[0063] An acquisition module is used to obtain historical long-term power load data;
[0064] The training module is used to train the CEEMD-LSTM model based on historical long-term power load data;
[0065] The prediction module is used to predict the power load size at a predetermined time in the future based on the trained CEEMD-LSTM model;
[0066] The training modules include:
[0067] A decomposition unit, used for decomposing the historical long-term power load data into multiple intrinsic mode function components through CEEMD;
[0068] A training unit, used for performing deep learning training on a plurality of long short-term memory neural network models respectively according to a plurality of intrinsic mode function components;
[0069] The prediction module includes:
[0070] A prediction unit, used to predict future scheduled time data through multiple trained long short-term memory neural network models to obtain multiple IMF component prediction values;
[0071] The reconstruction unit is used to reconstruct the prediction values of multiple IMF components to obtain the final prediction result.
[0072] Embodiment 3:
[0073] An embodiment of the present invention further provides a long-term power load forecasting system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a long-term power load forecasting method when executed by the processor.
[0074] Embodiment 4:
[0075] An embodiment of the present invention further provides a storage medium, on which a computer program is stored, and the computer program executes a long-term power load forecasting method when running.
[0076] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the protection scope determined by the claims of the present invention.
Claims
1. A long-term power load forecasting method, characterized in that: include: Step S1, obtaining historical long-term power load data; Step S2: training the CEEMD-LSTM model based on historical long-term power load data; Step S3: predicting the power load size at a predetermined time in the future according to the trained CEEMD-LSTM model; wherein, Step S2 includes: The historical long-term power load data is decomposed into multiple intrinsic mode function components through CEEMD; According to multiple intrinsic mode function components IMF, multiple long short-term memory neural network models are trained for deep learning respectively; Step S3 includes: Predict future scheduled time data through multiple trained long short-term memory neural network models to obtain multiple IMF component prediction values; Reconstruct the forecast values of multiple IMF components to obtain the final forecast result.
2. A long-term power load forecasting device, characterized in that: include: An acquisition module is used to obtain historical long-term power load data; The training module is used to train the CEEMD-LSTM model based on historical long-term power load data; The prediction module is used to predict the power load size at a predetermined time in the future based on the trained CEEMD-LSTM model; The training modules include: A decomposition unit, used for decomposing the historical long-term power load data into multiple intrinsic mode function components through CEEMD; A training unit, used for performing deep learning training on a plurality of long short-term memory neural network models respectively according to a plurality of intrinsic mode function components; The prediction module includes: A prediction unit, used to predict future scheduled time data through multiple trained long short-term memory neural network models to obtain multiple IMF component prediction values; The reconstruction unit is used to reconstruct the prediction values of multiple IMF components to obtain the final prediction result.
3. A long-term power load forecasting system, characterized in that: include: A memory and a processor, wherein the memory stores a computer program executed by the processor, and when the computer program is executed by the processor, the long-term power load forecasting method according to claim 1 is executed.
4. A storage medium, characterized in that: The storage medium stores a computer program, which executes the long-term power load forecasting method according to claim 1 when running.