Short-term power load prediction method and device based on weather similar days

Abstract

The invention discloses a short-term power load prediction method and device based on weather similar days, and the method comprises the steps: obtaining the load waveform data of a historical day, and dividing the waveform types; according to the meteorological factors and date categories of the to-be-tested day, determining a waveform category to which the to-be-tested day belongs by using a pre-trained classification model; based on meteorological factors of the to-be-measured day, the similarity of elements in the waveform category is judged, and the best similar day is obtained; and predicting the power load at the to-be-measured moment according to the load data at the to-be-measured moment corresponding to the optimal similar day and the meteorological factors at the to-be-measured moment. According to the method, the complex nonlinear relationship between the meteorological factors and the power load can be processed, so that the load prediction precision is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power load forecasting, and more particularly to a method and device for short-term power load forecasting based on meteorologically similar days. Background Art

[0002] In the field of power load forecasting, accurate prediction of short-term power load is crucial for the stable operation of the power grid and the effective management of energy. Traditional power load forecasting methods mainly rely on historical load data and simple statistical models, but these methods often ignore the impact of meteorological factors on power load, resulting in limited prediction accuracy. In recent years, with the development of big data and artificial intelligence technology, power load forecasting methods based on machine learning have gradually become a research hotspot. However, existing methods still have challenges in dealing with complex meteorological conditions and nonlinear relationships, especially in short-term forecasts. Therefore, how to make full use of the information of meteorologically similar days to improve prediction accuracy is an urgent problem to be solved. Summary of the invention

[0003] In view of this, in order to fully exploit the similar day information of different meteorological factors and the information contained in the input characteristics to improve the load forecasting accuracy, the present invention provides a short-term power load forecasting method and device based on meteorological similar days;

[0004] In order to achieve the above object, the present invention adopts the following technical solution:

[0005] In a first aspect, the present application provides a short-term power load forecasting method based on meteorologically similar days, comprising:

[0006] Obtain historical daily load waveform data and classify the waveforms to analyze the load change patterns of different types of days, providing a basis for subsequent similar day searches;

[0007] According to the meteorological factors and date category of the day to be tested, the waveform category is determined using the pre-trained classification model;

[0008] Based on the meteorological factors of the day to be measured, the similarity with the elements in the corresponding waveform category is judged to obtain the best similar day;

[0009] The power load at the time to be measured is predicted based on the load data of the best similar day corresponding to the time to be measured and the meteorological factors at the time to be measured.

[0010] Preferably, the meteorological factor is a meteorological factor that is strongly correlated with the load, and is determined by analyzing linear correlation and nonlinear correlation.

[0011] Preferably, the date categories refer to holidays and working days;

[0012] Preferably, the morphological similarity distance is used to judge the similarity with the elements in the corresponding waveform category, and similar day searches are preferentially performed within the range of the first m days of the same type before the day to be predicted, which helps to find historical days with meteorological conditions more similar to the day to be predicted, thereby improving the prediction accuracy.

[0013] Preferably, the power load at the time to be measured is predicted using a pre-trained load prediction model according to the load data at the time to be measured corresponding to the best similar day and the meteorological factors at the time to be measured;

[0014] The load forecasting model consists of TCN and BILSTM networks;

[0015] TCN network is used to extract and fuse the input data to obtain high-dimensional abstract features;

[0016] The BILSTM network is used to learn the complex interactions between time series to predict the power load at the time to be tested.

[0017] The load forecasting model of this application can better learn the complex interactive relationships between time series, thereby improving the accuracy and reliability of the forecast.

[0018] Preferably, before training the load forecasting model, the training data is preprocessed to ensure the quality of the training data and improve the stability and generalization ability of the model. The preprocessing includes:

[0019] The quartile method is used to detect outliers in the data, convert them into missing data, and fill in the data through the upsampling interpolation function;

[0020] The original data is converted to the range of [0,1] by using the linear function normalization method, and the predicted data is restored to the original scale by the anti-normalization equation after prediction.

[0021] Preferably, the power load at the time to be measured is predicted based on the load data at the time to be measured corresponding to the best similar day, the load data at the time to be measured n days before the day to be measured, and the meteorological factors at the time to be measured. This can more comprehensively capture the changing trend of short-term loads and further improve the prediction accuracy.

[0022] In a second aspect, the present application provides a short-term power load forecasting device based on meteorologically similar days, comprising:

[0023] A load waveform classification unit is used to obtain historical daily load waveform data and classify waveform categories;

[0024] A waveform category determination unit, used to determine the waveform category using a pre-trained classification model according to the meteorological factors and date category of the day to be measured;

[0025] A similar day determination unit is used to determine the similarity with the elements in the corresponding waveform category based on the meteorological factors of the day to be measured, and obtain the best similar day;

[0026] The power load prediction unit is used to predict the power load at the time to be tested based on the load data of the best similar day corresponding to the time to be tested and the meteorological factors at the time to be tested.

[0027] Optionally, the power load prediction unit is used to predict the power load at the time to be measured based on the load data at the time to be measured corresponding to the best similar day, the load data at the time to be measured n days before the day to be measured, and the meteorological factors at the time to be measured.

[0028] It can be seen from the above technical solutions that the present invention discloses a method and device for short-term power load forecasting based on meteorologically similar days. Compared with the prior art, the present invention can handle the complex nonlinear relationship between meteorological factors and power load, has strong adaptability, can cope with different weather conditions and load change patterns, and can effectively improve the load forecasting accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] 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.

[0030] Figure 1 This is a flow chart of a short-term power load forecasting method based on meteorologically similar days according to the present invention;

[0031] Figure 2 This is a multi-scenario load clustering result diagram of the present invention;

[0032] Figure 3 The Pearson correlation coefficient heat map of each season of the present invention;

[0033] Figure 4 This is the heat map of MIC correlation coefficients in each season of the present invention;

[0034] Figure 5 The load curve diagram of the best similar day of the present invention;

[0035] Figure 6 It is a schematic diagram of the structure of the load forecasting model of the present invention;

[0036] Figure 7 Schematic diagram of the residual unit structure of the present invention;

[0037] Figure 8 This is a comparison chart of prediction curves of different models of the present invention. DETAILED DESCRIPTION

[0038] 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.

[0039] The embodiment of the present invention discloses a method for short-term power load forecasting based on meteorological similarity days, the steps are as follows: Figure 1 , specifically including:

[0040] In one embodiment,

[0041] Step 1: Obtain historical daily load waveform data and classify the waveform categories;

[0042] The power load is affected by many factors, including seasons, weather, working days and holidays, and will show certain regularity and periodicity according to the similarities and differences of the influencing factors. In order to better analyze the trend of power load changes, the K-means clustering algorithm is used to perform cluster analysis on historical power load data. In this way, dates with similar load characteristics can be grouped to form similar day scenarios.

[0043] K-means clustering is an unsupervised clustering algorithm that aims to automatically divide N samples in a G-dimensional space into k categories according to their data characteristics, so that the similarity of objects within a class is large and the similarity of objects between classes is small. The clustering process of this algorithm is divided into four steps:

[0044] (1) Class center point initialization: Randomly select N samples {x1,x2,…,x N} as the initial cluster center points {t1, t2, …, t k}.

[0045] (2) Classification: The following formulas are used to calculate x1, x2, …, x N with t1,t2,…,t k The Euclidean distance between them and the center point with the smallest distance is classified into one category to form k clusters {T1, T2, …, T k}.

[0046]

[0047] (3) Cluster center solution: Calculate the average value of all samples in each category and update the cluster center.

[0048] (4) Convergence judgment: Determine whether convergence is achieved according to formula (2). If the requirements are not met, repeat steps (2) and (3) until the criterion function R reaches the set magnitude range, and clustering is completed.

[0049]

[0050] Among them, T i represents the i-th class sample set, x q T i Sample object, t i For the T i The cluster center of the class.

[0051] In one embodiment, cluster analysis is performed using the data from the entire year of 2016 in the data set, and the clustering results are shown in Figure 2 .Depend on Figure 2 It can be seen that after clustering the power load data using the K-means clustering algorithm, four different clustering results are obtained. Each graph represents a specific clustering category, representing the power load pattern under different temperature characteristics and date characteristics.

[0052] The first type of load curves are relatively concentrated, with a peak occurring around noon, indicating that this type of load is mainly in a period of time when the temperature is relatively stable, which may correspond to the load characteristics of weekdays in spring and autumn; the second type of load curve shows two peaks at noon and in the evening, and the distribution between the curves is relatively scattered, indicating that this type of load pattern may be greatly affected by fluctuations in temperature and date, such as high temperatures in summer or special holidays; the third type of load curve gradually rises in the early morning and reaches a peak in the evening, reflecting the more obvious characteristics of electricity demand in the morning and evening, which may be related to the heating demand in cold weather; the fourth type of curve shows a peak in the afternoon and decreases in the afternoon, showing that the load demand is high during the day but tends to be stable at night, which may be related to the peak of air conditioning use in hot weather.

[0053] Through clustering result analysis, days with similar load characteristics are identified, providing a basis for the construction of subsequent similar day selection mechanism.

[0054] In one embodiment,

[0055] Step 2: According to the meteorological factors and date categories of the day to be tested, use the pre-trained classification model to determine the waveform category; the date category refers to holidays and working days

[0056] Meteorological factors include temperature, humidity, and wind speed. Since the degree of influence of each factor on power load varies, it is necessary to accurately identify the key features that affect the power load data and select meteorological factors that are strongly correlated with the load.

[0057] In one embodiment, in order to effectively identify the sensitive features that affect the power load, the present application implements mining of the highly correlated sensitive features of the load sequence through linear correlation and nonlinear correlation analysis.

[0058] In this embodiment, the Pearson correlation coefficient is used for linear correlation analysis, and its calculation formula is:

[0059]

[0060] Where: r is the Pearson correlation coefficient between variables; is the mean value of variable F; is the average value of variable L; n is the data sample size. The larger |r| is, the stronger the correlation is. Generally speaking, |r|≥0.6 indicates a strong correlation, |r|∈[0.4,0.6) indicates a moderate correlation, |r|∈[0.2,0.4) indicates a weak correlation, and |r|∈[0,0.2) indicates a very weak correlation or no correlation.

[0061] Furthermore, the maximum information coefficient is used for nonlinear correlation analysis. In a data set D, the interval is divided along the coordinate axis of the data set to obtain x×y grids. The probability distribution of the binary data set in each grid is D|G, and I(D|G) represents the mutual information of D|G. The maximum standardized mutual information of each grid of the data set D constitutes the feature matrix M(D), and the matrix elements are as follows:

[0062]

[0063] MIC(D)=max xy＜B(n) M(D) x,y

[0064] Among them, the product of x and y must be less than B(n), generally B(n) = n 0.6 The MIC metric is generally to reflect the correlation between variables by comparing the size of the MIC value. The larger the MIC value, the stronger the correlation between the two variables.

[0065] In one embodiment, the power load and meteorological data of a region in 2016 and 2017 are analyzed, with a resolution of 1 hour, and three meteorological factors are mainly considered: temperature (degrees Celsius), humidity (%), and wind speed (m / s). n ] represents various meteorological factors, L=[L1,L2,…,L n ] represents the load, and the calculation results are shown in Table 1 and Table 2. The thermal diagram corresponds to Figure 3 and Figure 4 ;

[0066] Table 1

[0067] Meteorological factors spring summer autumn winter temperature 0.72 0.75 0.73 0.72 humidity -0.14 -0.01 0.13 -0.04 Wind speed 0.03 0.07 0.08 0.18

[0068] Table 2

[0069] Meteorological factors spring summer autumn winter temperature 0.43 0.46 0.44 0.48 humidity 0.05 0.04 0.04 0.03 Wind speed 0.02 0.02 0.02 0.03

[0070] As shown in Table 1, in spring, summer, autumn and winter, the temperature is greater than 0.7, indicating that there is a strong degree of correlation between load and temperature. Especially in summer and autumn, it reaches 0.75 and 0.73, indicating that temperature and load have the strongest correlation in summer and autumn. The spring and winter seasons also reach 0.72, slightly lower than the summer and autumn seasons, and also show a strong correlation; humidity and wind speed are both less than 0.2, indicating that there is a weak correlation between humidity, wind speed and load. Humidity in spring and autumn is greater than 0.1, generally speaking, there is a certain correlation but not large, while in summer and winter it is less than 0.1, which can be said to have a very weak correlation. Compared with humidity, wind speed generally has a weak correlation in spring, summer, autumn and winter, which can be ignored. Combining the analysis results of each season, load and temperature have a stronger linear correlation than wind speed and humidity.

[0071] As shown in Table 2, in spring, summer, autumn and winter, the MIC of temperature is much greater than the MIC of humidity and wind speed, indicating that the nonlinear correlation between load and temperature is stronger than that between humidity and wind speed. The MIC of temperature in spring, summer, autumn and winter is greater than 0.4, and the MIC in summer and winter reaches 0.46 and 0.48, which is greater than the MIC in spring and autumn. Humidity and wind speed are also meteorological characteristics related to load. The MIC of humidity and wind speed is less than 0.1, which is significantly lower than that of temperature. However, the overall MIC value of humidity in the four seasons is still greater than the MIC of wind speed, indicating that the nonlinear correlation between load and humidity is slightly stronger than that of wind speed. Combining the analysis results of each season, load and temperature have a stronger nonlinear correlation than wind speed and humidity.

[0072] In summary, according to the linear and nonlinear correlation analysis results, it is shown that load has a stronger correlation with temperature.

[0073] Furthermore, in order to determine the waveform category to which the day to be tested belongs, an incremental extreme learning machine (Ⅰ-ELM) classification model is built, and the daily attributes of load influencing factors such as meteorological factors and date types in historical data are used as input, and the category of daily load is used as output for training.

[0074] Then, according to the meteorological factors and date category of the day to be tested, the pre-trained classification model is used to determine the waveform category.

[0075] In this embodiment, the specific steps are as follows:

[0076] (1) Input data

[0077] Construct the input matrix X and output label Y. The input matrix X contains the feature combination of the two main daily attributes:

[0078]

[0079] Where A1 and A2 represent the meteorological factors and date types in the historical data, respectively.

[0080] The output label Y indicates the daily load category corresponding to the historical data:

[0081]

[0082] (2) Random initialization of weights and biases

[0083] The weight matrix W and bias vector b between the input layer and the hidden layer are randomly generated; the size of the weight matrix W from the input layer to the hidden layer is L×m, where L is the number of hidden layer neurons and m is the input feature dimension; the size of the bias vector b is L×1.

[0084] (3) Calculate the hidden layer output matrix H

[0085] The calculation formula of the hidden layer output matrix H is:

[0086] H=g(X·W T +b)

[0087] Where g(·) is the activation function and X is the input matrix.

[0088] (4) Calculate the output weight matrix β

[0089] The generalized inverse matrix is ​​used to solve the output weight matrix β, and the formula is as follows:

[0090] β=H -1 Y

[0091] (5) Model training

[0092] Use H and β to calculate the output of the training set, perform model training and update the weights.

[0093] (6) Prediction of test data

[0094] Enter the characteristic data X of the day to be tested pre , calculate the corresponding hidden layer output H:

[0095] H pre =g(X pre ·W T +b)

[0096] Predicted category Y pre Calculated by the following formula:

[0097] Y pre =H pre β

[0098] In one embodiment,

[0099] Step 3: Based on the meteorological factors of the day to be measured, determine the similarity with the elements in the corresponding waveform category and obtain the best similar day;

[0100] By building the I-ELM classification model and determining the category of the day to be predicted, the morphological similarity distance is used to judge the similarity with the elements in the corresponding waveform category, and the dates with closer distances are selected first to construct the meteorological similarity day set. If the best similarity day selected for temperature is the same day, the second similarity day of temperature and the best similarity day are taken to form the meteorological similarity day set.

[0101] Morphological similarity distance is a method for evaluating the similarity of sequence data. It can evaluate the similarity of sequence data from two perspectives: sequence data size and sequence data shape. The calculation formula is as follows:

[0102]

[0103] Where: L i is the real-time meteorological sequence of the day to be predicted, L j is the real-time meteorological sequence of historical days, l ik For the sequence L i The kth element of jk For the sequence L j The kth element of D Euclid is the Euclidean distance between two sequences, ASD is the absolute value of the sum of the differences between the values ​​of the two sequences, SAD is the Manhattan distance between the two sequences, D(L i ,L j ) is the morphological similarity distance between the two sequences. The historical day with the smallest morphological similarity distance is the best similarity day.

[0104] Taking into account the principle of “larger near and smaller far”, that is, the load is more relevant to the more recent information in the historical period, a search for similar days is performed within the range of the 10 days of the same type before the day to be predicted.

[0105] In one embodiment, January 1, 2017 is used as the forecast day to select the meteorological similar day, and the corresponding best similar day is selected using the real-time temperature sequence. The corresponding dates are December 19, 2016, December 25, 2016, and December 31, 2016. The load curve of the best similar day is as follows: Figure 5 shown.

[0106] Depend on Figure 5It can be seen that from the load curve shape, the load curve on December 19, 2016 is closest to the load curve on January 1, 2017, the day to be predicted. The distance between the load sequence of each best similar day and the load sequence of the day to be predicted is further calculated by the morphological similarity distance, and the selection results are evaluated, as shown in Table 3.

[0107] Table 3

[0108]

[0109] In the process of selecting similar days, the morphological similarity distance of December 19, 2016 was calculated to be 0.6631, which shows that it is most similar to January 1, 2017. This shows that the temperature sequence of December 19, 2016 is closer to the temperature change trend of the predicted day and is the best similar day. By using a single temperature sequence for similar day selection, the interference of other meteorological factors can be effectively avoided, and more accurate similar day results can be obtained.

[0110] In one embodiment,

[0111] Step 4: Predict the power load at the time to be tested using a pre-trained load prediction model based on the load data at the time to be tested corresponding to the best similar day and the meteorological factors at the time to be tested.

[0112] Since TCN has an extended causal convolution structure and outstanding feature extraction capabilities, it can fuse the original features to obtain high-dimensional abstract features, thereby strengthening the mining of feature information. The BILSTM network has a strong time series prediction capability. Combining TCN and BILSTM networks, extracting TCN features and inputting them into the BILSTM network improves the processing efficiency of the BILSTM network memory unit, allowing the prediction model to more effectively learn the complex interactive relationships of time series.

[0113] Therefore, the load forecasting model of this application consists of TCN and BILSTM networks;

[0114] TCN network is used to extract and fuse the input data to obtain high-dimensional abstract features;

[0115] The BILSTM network is used to learn the complex interactions between time series to predict the power load at the time to be tested.

[0116] The load forecasting model of this application can better learn the complex interactive relationships between time series, thereby improving the accuracy and reliability of the forecast.

[0117] Specific structure reference Figure 6 ,include:

[0118] 1) The input layer is the power load sequence and characteristics;

[0119] 2) In the temporal convolutional network layer TCN, a 1-layer residual unit is used. The residual unit structure is as follows: Figure 7 As shown, in this embodiment, the residual unit is composed of two causal convolution units and multiple nonlinear mappings.

[0120] Adding causal convolution in TCN can ensure that there is no information leakage from the future to the past. The application of DCC can have a larger field of view even if TCN has fewer layers, which means it can receive longer time series. ReLu activation function, weight normalization, Dropout regularization and other operations are used in DCC.

[0121] The residual connection adds the input x to the output f(x). f(x) can be expressed as:

[0122] f(x)=h(x)-x

[0123] 3) The output vector of the TCN model is the input of BILSTM;

[0124] 4) A fully connected layer is set in the output layer, and the output is the load forecast value at one moment.

[0125] In order to further optimize the above technical solution and solve the problems of missing data and abnormal fluctuations in the original power load data, this application adopts the quartile method to detect outliers in the data, converts them into missing data, and fills the data through an upsampling interpolation function; to prevent the problem of dimensionality explosion during the training process.

[0126] Since the data dimensions of load and meteorology are different, in order to avoid the impact of different data dimensions on model training, the linear function normalization method is used to convert the original data to the range of [0,1] to improve the efficiency and stability of model training. The normalization equation is as follows:

[0127]

[0128] Where, X min and X max are the minimum and maximum values ​​of the input data matrix X. The normalized data X norm Making the input features on the same scale helps speed up the convergence of the model.

[0129] In order to restore the normalized data to the original scale after prediction, the denormalization equation is used as follows:

[0130] X restore =X norm ·(X max -X min )+X min

[0131] In order to further optimize the above technical solution, this application:

[0132] (1) Considering the influence of meteorological similar days, the load value at time t on the day with the best temperature similarity corresponding to the predicted day is selected as the meteorological similar day feature.

[0133] (2) Considering that the load curve has daily cycle characteristics and autoregressive characteristics, the load at time t three days before the predicted day is selected as the historical load feature.

[0134] (3) Considering the impact of meteorological factors on load, according to the correlation analysis results between load and meteorological factors, temperature has a stronger correlation with load. Therefore, the temperature at time t is selected as the meteorological feature.

[0135] That is, the input features are shown in Table 4;

[0136] Table 4

[0137]

[0138]

[0139] In order to meet the data input requirements of the model, a sliding time window is used to divide the data, where the time step is set to 32, and the data input is shown in the following formula:

[0140]

[0141] Where, X t is the input of the sliding time window at time t, m is the number of input features, X is the input at the prediction time, and Y is the output at the prediction time.

[0142] In another embodiment, the present application provides a short-term power load forecasting device based on meteorologically similar days, comprising:

[0143] A load waveform classification unit is used to obtain historical daily load waveform data and classify waveform categories;

[0144] A waveform category determination unit, used to determine the waveform category using a pre-trained classification model according to the meteorological factors and date category of the day to be measured;

[0145] A similar day determination unit is used to determine the similarity with the elements in the corresponding waveform category based on the meteorological factors of the day to be measured, and obtain the best similar day;

[0146] The power load prediction unit is used to predict the power load at the time to be tested based on the load data of the best similar day corresponding to the time to be tested and the meteorological factors at the time to be tested.

[0147] Optionally, the power load prediction unit is used to predict the power load at the time to be measured based on the load data at the time to be measured corresponding to the best similar day, the load data at the time to be measured n days before the day to be measured, and the meteorological factors at the time to be measured.

[0148] In this embodiment, the execution process of each unit in the prediction device is consistent with the above-mentioned prediction method, so it will not be repeated here.

[0149] Furthermore, in order to verify the effectiveness of the proposed method, the power load and meteorological data of a certain region in 2016 and 2017 were used for analysis and calculation. The sampling interval of the data was 1 hour, that is, 24 time points were sampled every day, and three meteorological factors were mainly considered: temperature (degrees Celsius), humidity (%), and wind speed (m / s).

[0150] The laboratory hardware configuration is a computer with an Intel Core i9-13900HX CPU and 16GB of memory, accelerated by NVIDIA GeForce RTX 4070 GPU, and the experiment is programmed using Matlab 2023a and PyCharm.

[0151] In this embodiment, in order to obtain the optimal structure of the prediction model, the number of neurons in the BILSTM layer is set to 60, the maximum number of iterations (epoch) is set to 100 times, and the activation function is a linear rectifier function (ReLU). In order to avoid overfitting of the model, a Dropout layer with a deactivation rate of 0.02 is added. The initial learning rate is set to 0.005, and the learning rate is halved after every 2 epochs. The learning rate is automatically adjusted by the callback function, and other parameters are determined by repeated adjustment and optimization. The number of filters in the convolutional layer is set to 10, and the filter size is 2 to ensure that the model can effectively extract local features in time series data.

[0152] A residual block was used to alleviate the gradient vanishing problem and ensure the training stability of the deep network. During the training process, the Adam optimizer was selected and the above parameter settings were determined through repeated experiments to improve the prediction accuracy and generalization ability of the model.

[0153] In order to evaluate the prediction accuracy of the prediction model, the mean absolute error (MAE), root mean square error (RMSE) and goodness of fit R were used. 2 As an evaluation index, the calculation formula is as follows:

[0154]

[0155] In the formula, represents the predicted value of power load at the i-th moment; y i represents the actual value of the power load at the i-th moment; Represents the predicted average value of power load. The smaller the MAE and RMSE, the better the model prediction effect. 2 The closer it is to 1, the better the prediction model fits.

[0156] Furthermore, the original load sequence and feature set are used as the input of the prediction model. A total of 17,544 sets of data from January 1, 2016 to December 31, 2017 are predicted using TCN-BILSTM, BILTM and LSTM prediction models, which are divided into training set and test set in a ratio of 5:1. In order to intuitively understand the prediction results, the prediction results of three consecutive days are randomly selected from the test set for display and analysis. The experimental results are shown in Figure 2. Figure 8 shown.

[0157] As can be seen in the figure, the TCN-BILSTM model shows good stability in short-term power load forecasting, which is basically consistent with the actual load value. The model can effectively grasp the overall trend of load changes, especially in the period of gentle load fluctuations. The overlap between the predicted curve and the actual curve is high, indicating that the model has a high fitting accuracy within the normal load fluctuation range. In addition, the model prediction output is relatively smooth, without obvious overfitting, especially in extreme points and rapidly changing load areas, it can effectively grasp the load changes, and the prediction accuracy of peak load and valley load is relatively high. The error comparison of load forecast results is shown in Table 5.

[0158] Table 5

[0159] Model MAE / MW RMSE / MW <![CDATA[R 2 / %]]> TCN-BILSTM 229.49 297.48 0.97 BILSTM 263.87 327.60 0.96 LSTM 431.82 538.85 0.91

[0160] As shown in Table 5, the TCN-BILSTM model is significantly better than the other two models in terms of MAE and RMSE indicators. Specifically, the MAE indicator of TCN-BILSTM is reduced by 13.3% compared with BILSTM and 46.86% compared with LSTM. The RMSE indicator is reduced by 9.19% compared with BILSTM and 44.79% compared with LSTM. In addition, the goodness of fit of TCN-BILSTM reaches 97%, which is higher than BILSTM and LSTM, indicating that its performance and stability in power load forecasting have obvious advantages.

[0161] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0162] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A short-term power load forecasting method based on meteorological similarity days, characterized in that: Obtain historical daily load waveform data and divide the waveform categories; According to the meteorological factors and date category of the day to be tested, the waveform category is determined using the pre-trained classification model; Based on the meteorological factors of the day to be measured, the similarity with the elements in the corresponding waveform category is judged to obtain the best similar day; The power load at the time to be measured is predicted based on the load data of the best similar day corresponding to the time to be measured and the meteorological factors at the time to be measured.

2. The short-term power load forecasting method according to claim 1, characterized in that: The meteorological factors are meteorological factors that are strongly correlated with the load and are determined by analyzing the linear correlation and the nonlinear correlation.

3. The short-term power load forecasting method according to claim 1, characterized in that: Date categories refer to holidays and working days.

4. The short-term power load forecasting method according to claim 1, characterized in that: The morphological similarity distance is used to judge the similarity with the elements in the corresponding waveform category, and the search for similar days is performed preferentially within the range of the first m days of the same type before the day to be predicted.

5. The short-term power load forecasting method according to claim 1, characterized in that: According to the load data of the best similar day corresponding to the time to be tested and the meteorological factors at the time to be tested, the power load at the time to be tested is predicted using the pre-trained load prediction model; The load forecasting model consists of TCN and BILSTM networks; TCN network is used to extract and fuse the input data to obtain high-dimensional abstract features; The BILSTM network is used to learn the complex interactions between time series to predict the power load at the time to be tested.

6. The short-term power load forecasting method according to claim 4, characterized in that: Before training the load forecasting model, the training data is preprocessed, including: The quartile method is used to detect outliers in the data, convert them into missing data, and fill in the data through the upsampling interpolation function; The original data is converted to the range of [0,1] by using the linear function normalization method, and the predicted data is restored to the original scale by the anti-normalization equation after prediction.

7. The short-term power load forecasting method according to claim 1, characterized in that: The power load at the time to be measured is predicted based on the load data at the time to be measured corresponding to the best similar day, the load data at the time to be measured n days before the day to be measured, and the meteorological factors at the time to be measured.

8. A short-term power load forecasting device based on meteorological similarity days, characterized in that: include: A load waveform classification unit is used to obtain historical daily load waveform data and classify waveform categories; A waveform category determination unit, used to determine the waveform category using a pre-trained classification model according to the meteorological factors and date category of the day to be measured; A similar day determination unit is used to determine the similarity with the elements in the corresponding waveform category based on the meteorological factors of the day to be measured, and obtain the best similar day; The power load prediction unit is used to predict the power load at the time to be tested based on the load data of the best similar day corresponding to the time to be tested and the meteorological factors at the time to be tested.

9. The short-term power load forecasting device according to claim 8, characterized in that: The power load prediction unit is used to predict the power load at the time to be tested based on the load data at the time to be tested corresponding to the best similar day, the load data at the time to be tested n days before the day to be tested, and the meteorological factors at the time to be tested.

Images