Short-term power load prediction method in combination with similar day factors

Abstract

The invention discloses a short-term power load prediction method combined with similar day factors, and the method comprises the steps: S1, carrying out the normalization processing of meteorological features, building an initial data set based on the processed meteorological features, holiday and festival features and load data, and carrying out the clustering and correlation analysis of the initial data set, and obtaining a similar day data set; s2, constructing a CNN-GRU model, training the CNN-GRU model through a similar day data set, and adjusting model parameters through a grid search method; s3, stopping model training when verification set parameter searching is completed, evaluating the trained model, searching for a next group of parameters until parameter searching is completed, comparing model effects under the parameters, and selecting an optimal model; and S4, predicting the short-term power load based on the optimal model.

Description

Technical Field

[0001] The present invention relates to the technical field of short-term power load forecasting, and more particularly to a short-term power load forecasting method combining similar day factors. Background Art

[0002] The method of short-term power grid load forecasting is based on the historical load data of the region. In the forecasting process, the influence of various external factors is considered at the same time. By establishing a mathematical relationship to reflect the implicit complex relationship between the two, the forecast value of the future power load in the region is obtained. Improving the accuracy of short-term load forecasting can not only improve the safe and stable operation of the power grid, but also ensure reliable power supply.

[0003] The most significant feature of short-term load forecasting is its obvious periodicity, including:

[0004] (1) Similarity of the overall 24-hour variation patterns between different days.

[0005] (2) Similarities between types of days in different weeks and in the same week.

[0006] (3) Similarities between working days and rest days.

[0007] (4) Similarity of load curves during major holidays in different years.

[0008] In addition to the above-mentioned periodicity, another characteristic of short-term load is that it is obviously affected by various environmental factors, such as seasonal changes, sudden changes in weather factors, equipment accidents and maintenance, major problem activities, etc., which makes the changes in the load event sequence appear to be a non-stationary random process.

[0009] At present, the short-term load forecasting method currently used by the power grid in the region is a traditional method based on traditional experience. It has a low accuracy rate and cannot well combine the load characteristics of similar days (similar weather days, similar major festival days), and cannot meet the current development needs of power grid load forecasting.

[0010] The publication number is CN111382891A, and the name is short-term load forecasting method and short-term load forecasting device. The short-term load forecasting method includes: obtaining load data of the past period and external parameters of the forecast period; decomposing the load data of the past period into second-level detail components, third-level approximate components and third-level detail components by wavelet decomposition; obtaining the second-level detail components of the forecast load data of the forecast period based on the second-level detail components; obtaining the third-level approximate components and third-level detail components of the forecast load data of the forecast period based on the third-level approximate components and third-level detail components of the load data of the past period and the external parameters; reconstructing the second-level detail components, the third-level approximate components and the third-level detail components of the forecast load data into the forecast load data of the forecast period by wavelet reconstruction. According to the short-term load forecasting method and the short-term load forecasting device, short-term load can be accurately forecasted.

[0011] The publication number is CN116404637A, and the name is "Power System Short-term Load Forecasting Method and Power System Short-term Load Forecasting Device". It discloses a short-term load history curve based on each prediction scenario, combined with the time series and distribution similarity of each short-term load history curve for clustering, to obtain the optimal clustering result; according to different target application scenarios, different load forecasting models based on Transformer are established, and the short-term load history curve and known features are used to train and evaluate the load forecasting model in each category to obtain the best performance model for each category; the best performance model for each category is migrated, so that each best performance model is applied to other application scenarios between categories for short-term load forecasting. The present invention concisely performs short-term load forecasting for any application scenario. It improves the accuracy of low-level load forecasting, with small amount of calculation and fast calculation speed.

[0012] Holiday load forecasting has the following characteristics: The load change pattern during holidays is significantly different from that on normal days. Compared with normal working days, the load on holidays is generally significantly lower. For example, during the Spring Festival, the load curve will show a long-term, large-scale decline and deformation. Since there is less load data during holidays, the load is easily affected by various random fluctuation factors and potential interference factors, making it difficult for some forecasting algorithms to obtain relatively satisfactory results.

[0013] Therefore, how to provide a more accurate short-term load forecasting method that conforms to the local load characteristics is an urgent problem that technicians in this field need to solve. Summary of the invention

[0014] In view of this, the present invention provides a short-term power load forecasting method combining similar day factors, aiming to solve the above technical problems.

[0015] In order to achieve the above object, the present invention adopts the following technical solution:

[0016] A short-term power load forecasting method combining similar day factors, including:

[0017] Step S1, normalizing the meteorological characteristics, constructing an initial data set based on the processed meteorological characteristics, holiday characteristics and load data, performing clustering and association analysis on the initial data set, and obtaining a similar day data set;

[0018] Step S2, construct a CNN-GRU model, train it with similar day data sets, and adjust model parameters by grid search method;

[0019] Step S3: stop model training when the validation set parameter search is completed, evaluate the trained model, search for the next set of parameters, and compare the model effects under various parameters to select the optimal model.

[0020] Step S4: forecast short-term power load based on the optimal model.

[0021] Furthermore, the meteorological data include average wind speed x1, accumulated precipitation x2, temperature x3, accumulated sunshine x4 and average humidity x5, and the load data Y is the maximum load value of the day.

[0022] Furthermore, in step S1, the meteorological data is normalized as follows:

[0023] Based on the MinMaxScaler function, the meteorological data is normalized so that the meteorological data is located in [0,1], specifically:

[0024]

[0025] Among them, x i For meteorological data.

[0026] Furthermore, in step S1, the initial data is clustered, specifically:

[0027] Step S11, randomly selecting a sample from the initial data as the initial cluster center;

[0028] Step S12: Calculate the Euclidean distance between each sample in the initial data and the initial cluster center, and calculate the probability θ(x) of each sample being selected as the next cluster center according to the following formula:

[0029]

[0030] Among them, x is the sample, is the sum of the squares of all individual distance measures in set X;

[0031] Select the next cluster center based on the roulette wheel method;

[0032] Step S13, repeat step S12 until K cluster centers are selected;

[0033] Step S14: for each sample in the initial data, calculate the distance from each sample to the K cluster centers, and divide each sample into the class corresponding to the cluster center with the smallest distance;

[0034] Step S15: Calculate the cluster center C of each class i

[0035]

[0036] Step S16: Repeat steps S14-S15 until the location of the cluster center remains unchanged and a rough set is obtained.

[0037] Furthermore, the correlation analysis in step S1 is performed to obtain a similar day data set, specifically:

[0038] Step S16: Confirm the factors affecting the power load, the total sequence of factors affecting the i-th day X i for:

[0039] X i =(x 1i ,x 2i ,...,x ni )

[0040] Among them, n is the characteristic value of similar day factor, i=1,2,…,m, m is the number of days of rough set records, and the characteristic subsequence X0 of the day to be predicted is selected as:

[0041] X0=(x 10 ,x 20 ,...,x n0 )

[0042] Step S17: Divide each element in the total sequence and the subsequence by the average value in the respective vectors for averaging. Let the vector after averaging be X′ i and X′0, X′ i The relationship coefficient with X′0 is ξ i(k) :

[0043]

[0044] Where k = 1, 2, …, n, |X′ 0(k) -X′ i(k) | is the absolute difference between the total sequence and the subsequence influencing factors after averaging, and ρ is the resolution coefficient;

[0045] Step S18: Each factor corresponds to a correlation coefficient, so there are n correlation coefficients, and their average value is the correlation degree r i :

[0046]

[0047] The data with correlation greater than the preset value are selected to form a data set similar to the day to be predicted.

[0048] Furthermore, the training process of the CNN-GRU model in step S2 is:

[0049] Initialize the CNN-GRU model, calculate through forward propagation, get the initial prediction result, calculate the error between the predicted value and the true load value, and judge whether it converges. If it converges, end the training. If not, backpropagate the error to the model and update the weights of the CNN and GRU models.

[0050] Furthermore, the weight update process of the CNN model is:

[0051] Use k one-dimensional convolution kernels to convolve the multivariate time series x1,x2,...,x T ∈R k The time series of each variable in Extract sequence features to obtain the feature vector of the i-th sequence The calculation process is shown in the formula:

[0052]

[0053] represents the jth component of the cth convolution kernel, b is the bias parameter, and f is the activation transformation function.

[0054] Furthermore, the weight update process of the GRU model is:

[0055]

[0056] Among them, u and r represent the update gate and the reset gate.

[0057] Furthermore, the grid search method in step S2 adjusts the model parameters specifically as follows:

[0058] Build a dictionary of parameters to be adjusted, which contains all the parameters that need to be adjusted in the model and their selectable values; define a classifier, put the parameters that need to be adjusted into the classifier, and get an object clf; build a GridSearchCV object; train the GridSearchCV object and find the optimal parameters.

[0059] Furthermore, the specific process of model evaluation is as follows:

[0060] Mean absolute percentage error:

[0061]

[0062] Root Mean Square Error:

[0063]

[0064] The above technical solution includes at least the following technical effects:

[0065] Through the above technical solutions, it can be known that the present invention discloses a short-term power load forecasting method combined with similar day factors. Compared with the prior art, the present invention introduces similar day features, uses training set data to train the model, and uses validation set data to verify the effect of the model. If the model performs poorly on the validation set, the model parameters need to be adjusted. Parameters include learning rate, number of nodes, number of hidden layers, etc. Generally, the more complex the model, the more parameters need to be adjusted. The model parameters are continuously adjusted until the parameter search is completed, and the best set of parameters is selected based on the performance of the model on the validation set. Finally, the model under the parameters is used to make a final prediction on the test set. . BRIEF DESCRIPTION OF THE DRAWINGS

[0066] 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.

[0067] Figure 1 This is a schematic diagram of the data set provided by the present invention.

[0068] Figure 2 This is a schematic diagram of the CNN-GRU model training process provided by the present invention. DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present invention are described clearly and completely below. 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.

[0070] See attached Figure 1-2 The embodiment of the present invention discloses a short-term power load forecasting method combining similar day factors, including:

[0071] Step S1, normalizing the meteorological characteristics, constructing an initial data set based on the processed meteorological characteristics, holiday characteristics and load data, performing clustering and association analysis on the initial data set, and obtaining a similar day data set;

[0072] Step S2, construct a CNN-GRU model, train it with similar day data sets, and adjust model parameters by grid search method;

[0073] Step S3: stop model training when the validation set parameter search is completed, evaluate the trained model, search for the next set of parameters, and compare the model effects under various parameters to select the optimal model.

[0074] Step S4: forecast short-term power load based on the optimal model.

[0075] To further optimize the above technical solution, the meteorological data include average wind speed x1, accumulated precipitation x2, temperature x3, accumulated sunshine x4 and average humidity x5, and the load data Y is the maximum load value of the day.

[0076] To further optimize the above technical solution, the meteorological data is normalized in step S1 as follows:

[0077] Based on the MinMaxScaler function, the meteorological data is normalized so that the meteorological data is located in [0,1], specifically:

[0078]

[0079] Among them, x i For meteorological data.

[0080] To further optimize the above technical solution, the initial data is clustered in step S1, specifically:

[0081] Step S11, randomly selecting a sample from the initial data as the initial cluster center;

[0082] Step S12: Calculate the Euclidean distance between each sample in the initial data and the initial cluster center, and calculate the probability θ(x) of each sample being selected as the next cluster center according to the following formula:

[0083]

[0084] Among them, x is the sample, is the sum of the squares of all individual distance measures in set X;

[0085] Select the next cluster center based on the roulette wheel method;

[0086] Step S13, repeat step S12 until K cluster centers are selected;

[0087] Step S14: for each sample in the initial data, calculate the distance from each sample to the K cluster centers, and divide each sample into the class corresponding to the cluster center with the smallest distance;

[0088] Step S15: Calculate the cluster center C of each class i

[0089]

[0090] Step S16: Repeat steps S14-S15 until the location of the cluster center remains unchanged and a rough set is obtained.

[0091] To further optimize the above technical solution, the association analysis in step S1 is performed to obtain a similar day data set, specifically:

[0092] Step S16: Confirm the factors affecting the power load, the total sequence of factors affecting the i-th day X i for:

[0093] X i =(x 1i ,x 2i ,...,x ni )

[0094] Among them, n is the characteristic value of similar day factor, i=1,2,…,m, m is the number of days of rough set records, and the characteristic subsequence X0 of the day to be predicted is selected as:

[0095] X0=(x 10 ,x 20 ,...,x n0 )

[0096] Step S17: Divide each element in the total sequence and the subsequence by the average value in the respective vectors for averaging. Let the vector after averaging be X′ i and X′0, X′ i The relationship coefficient with X′0 is ξ i(k) :

[0097]

[0098] Where k = 1, 2, …, n, |X′ 0(k) -X′ i(k) | is the absolute difference between the total sequence and the subsequence influencing factors after averaging, and ρ is the resolution coefficient;

[0099] Step S18: Each factor corresponds to a correlation coefficient, so there are n correlation coefficients, and their average value is the correlation degree r i :

[0100]

[0101] The data with correlation greater than the preset value are selected to form a data set similar to the day to be predicted.

[0102] To further optimize the above technical solution, the training process of the CNN-GRU model in step S2 is:

[0103] Initialize the CNN-GRU model, calculate through forward propagation, get the initial prediction result, calculate the error between the predicted value and the true load value, and judge whether it converges. If it converges, end the training. If not, backpropagate the error to the model and update the weights of the CNN and GRU models.

[0104] To further optimize the above technical solution, the weight update process of the CNN model is:

[0105] Use k one-dimensional convolution kernels to convolve the multivariate time series x1,x2,...,x T ∈R k The time series of each variable in Extract sequence features to obtain the feature vector of the i-th sequence The calculation process is shown in the formula:

[0106]

[0107] represents the jth component of the cth convolution kernel, b is the bias parameter, and f is the activation transformation function.

[0108] To further optimize the above technical solution, the weight update process of the GRU model is:

[0109]

[0110] Among them, u and r represent the update gate and the reset gate.

[0111] To further optimize the above technical solution, the grid search method in step S2 adjusts the model parameters as follows:

[0112] Build a dictionary of parameters to be adjusted, which contains all the parameters that need to be adjusted in the model and their selectable values; define a classifier, put the parameters that need to be adjusted into the classifier, and get an object clf; build a GridSearchCV object; train the GridSearchCV object and find the optimal parameters.

[0113] To further optimize the above technical aspects, the specific process of model evaluation is as follows:

[0114] Mean absolute percentage error:

[0115]

[0116] Root Mean Square Error:

[0117]

[0118] Example 1, load data comes from the daily maximum load of the local power grid from 2015 to 2020. Date data is described by "whether it is a holiday". Meteorological data comes from data from approximately 820 stations provided by the China Meteorological Data Network. After manual sorting, the meteorological data of Baoding area from 2015 to 2020 was obtained. The daily meteorological data is described by five features: "average wind speed", "cumulative precipitation", "temperature", "cumulative sunshine" and "average humidity". The meteorological data is normalized (standardized) using the MinMaxScaler function so that the data is between [0,1]. The formula is as follows:

[0119]

[0120] The above three types of data together form a 2191×7 data set A (2191: 2191 days from 2015 to 2020, 7 includes 5 meteorological features, 1 holiday feature and 1 load data). Its features are used as feature sequences for clustering and association analysis, and data with a correlation greater than 0.7 are selected to form a data set B similar to the day to be predicted.

[0121] The present invention first uses the K-means++ clustering algorithm to divide the original collected data set into a coarse set, and screens the predicted day clusters into a coarse set according to the Euclidean distance. Then, the grey correlation analysis method is used to analyze the correlation degree of the sample data in the coarse set, and the similar day data set is selected according to the correlation degree.

[0122] K-means++ is an improved K-means clustering algorithm, the difference lies in the method of selecting the initial cluster center. It has the clustering advantages of the K-means algorithm and avoids the disadvantage of sometimes finding weak clusters, achieving the purpose of improving the initial centroid selection and effectively improving the clustering effect. Compared with the K-means algorithm, the first cluster center of K-means++ is randomly selected, and then the next cluster center is selected based on the distance, and the sample point farthest from the previous cluster center is selected as the next cluster center. This method of selecting cluster centers can avoid the initial cluster centers being too concentrated or dispersed. The specific steps are as follows:

[0123] 1. Randomly select a point as the initial cluster center;

[0124] 2. Calculate the Euclidean distance D(x) between each point in the sample and its nearest initial cluster center, calculate the probability of each sample being selected as the next cluster center according to formula (1), and select the next cluster center according to the roulette wheel method:

[0125]

[0126] θ(x) represents the probability value associated with x; Represents the sum of the squares of the distance measures of all individuals in set X. This summation term is used to normalize the probability value to ensure that the sum of the probability values ​​of all individuals is 1

[0127] 3. Repeat step 2 until a total of K cluster centers are selected;

[0128] 4. For each sample in the data set, calculate its distance to the K cluster centers, and divide each sample into the class corresponding to the cluster center with the smallest distance;

[0129] 5. Calculate the cluster center C of each class according to formula (2) i :

[0130]

[0131] The above formula represents the formula for calculating the cluster center of the cluster. By calculating the weighted average position of all data points, the position of the cluster center can be obtained.

[0132] 6. Repeat steps 4 and 5 until the cluster center position remains unchanged

[0133] The present invention uses the silhouette coefficient S to measure the clustering effect. It evaluates the quality of the clustering result by calculating the compactness of samples within the cluster and the separation of samples between clusters. The closer S is to 1, the better the clustering effect. The calculation formula of the silhouette coefficient S is:

[0134]

[0135] Where a(i) represents the average distance between sample i and other samples in the same cluster; b(i) represents the average distance between sample i and the samples in the nearest cluster in different clusters, that is, the average distance from sample i to the nearest cluster.

[0136] Grey correlation is used to measure the degree of similarity or difference between trends of factors. The correlation shows the degree of similarity of changes of different factors over time. When determining the factors affecting power load, the total sequence of factors on the i-th day can be expressed as:

[0137] X i =(x 1i ,x 2i ,...,x ni )

[0138] In the above formula: n is the characteristic value of the similar day factor, in the present invention n = 6; i = 1, 2, ..., m, m is the number of days recorded in the rough set. Select the characteristic subsequence of the day to be predicted as X0, which can be expressed as:

[0139] X0=(x 10 ,x 20 ,...,x n0 )

[0140] Divide each element in the total sequence and subsequence by the average value in their respective vectors for averaging. Let the vector after averaging be X′ i and X′0. X′ i The relationship coefficient with X′0 is ξ i(k) :

[0141]

[0142] Where: K = 1, 2, ..., n; |X′ 0(k) -X′ i(k) | is the absolute difference between the total sequence and the subsequence influencing factors after averaging; ρ is the resolution coefficient, which is usually 0.5.

[0143] Each factor corresponds to a correlation coefficient, so there are n correlation coefficients, and their average value is the correlation degree r between the systems. i :

[0144]

[0145] The greater the correlation, the higher the correlation.

[0146] Example 2, load forecasting model based on CNN-GRU:

[0147] Convolutional Neural Networks (CNN)

[0148] Convolutional neural network is a deep neural network with convolution structure inspired by visual neurons. It has been widely used in image processing and natural language processing. The local connection and weight sharing of CNN can reduce the number of weights and thus reduce the difficulty of network optimization, which effectively reduces the complexity of the network and the occurrence of overfitting problems. Convolutional neural network has richer layers than neural network. It mainly includes feedforward neural network composed of cross-stacked convolutional layer, pooling layer and fully connected layer. It has the following three characteristics in structure: local continuity, aggregation and weight sharing. These characteristics can make convolutional neural network produce a certain degree of translation, compression and rotation invariance. Unlike the traditional feedforward neural network model, the calculation between convolutional layers has fewer parameters than neural network. The model obtains effective information units by using convolutional layer and pooling layer, and can automatically obtain feature vectors in data analysis. It can automatically obtain some feature vectors of data, reduce the complexity of calculation and improve the accuracy of processing.

[0149] Use k one-dimensional convolution kernels to convolve the multivariate time series x1,x2,...,x T ∈R k The time series of each variable in Extract sequence features to obtain the feature vector of the i-th sequence The calculation process is shown in the formula:

[0150]

[0151] represents the jth component of the cth convolution kernel, b is the bias parameter, and f is the activation transformation function.

[0152] Gated Recurrent Unit (GRU)

[0153] Since the three gated units in LSTM have different effects on improving its learning ability, the gates with less effect and their corresponding weights are omitted in GRU, that is, GRU is a simplified model of LSTM. GRU has fewer parameters and faster training speed. After simplifying the neural network structure, the learning efficiency of the recurrent neural network has been improved. The gated recurrent unit network corresponds to two units: the update gate unit and the reset gate unit. Among them, the reset gate and the input gate of the LSTM unit have similar functions. The update gate can not only complete the task of the forget gate, but also complete the task of the output gate.

[0154] The update method of GRU can be expressed as follows:

[0155]

[0156] CNN-GRU is a hybrid neural network model that combines CNN and GRU in series, making full use of the advantages of the two models. CNN is used as an extractor of input features to extract hidden information of input features and exclude redundant features. Then the GRU model is used to extract the time dependency of input data. The combination of the two models can respond to the timing of loads to improve the prediction accuracy of the load forecasting model.

[0157] The training method of the model is back propagation, that is, reverse calculation from the fully connected layer to the convolutional layer. The back propagation process is shown in the figure below. First, the CNN and GRU models are initialized, and then forward propagation is performed to obtain preliminary prediction results. The error between the predicted value and the actual load value is calculated. If it does not converge, the weights of GRU and CNN are updated through back propagation.

[0158] To evaluate the generalization ability of the model, the model is trained using the training set data, and the model effect is verified using the validation set data. If the model performs poorly on the validation set, the model parameters need to be adjusted. Parameters include learning rate, number of nodes, number of hidden layers, etc. Generally, the more complex the model, the more parameters need to be adjusted. The model parameters are adjusted continuously until the parameter search is completed, and the best set of parameters is selected based on the performance of the model on the validation set. Finally, the model under this parameter is used to make the final prediction on the test set.

[0159] The model process is as follows:

[0160] 1. Use a similar set as a data set, 70% as a training set, 20% as a validation set, and 10% as a test set.

[0161] 2. Send the training set data into the CNN-GRU model and use the grid search method to continuously adjust the model parameters.

[0162] 3. When the model parameter search in the validation set is completed, stop model training, use conventional evaluation indicators such as MAPE and RMSE to evaluate and record the trained model, search for the next set of parameters until the parameter search is completed, compare the model effects under each parameter, and select the optimal model.

[0163] 4. Input the test set data into the optimal model for short-term load forecasting and compare the prediction accuracy with other models.

[0164] Example 3: The present invention uses the following two indicators to evaluate the CNN-GRU model.

[0165] Mean Absolute Percentage Error (MAPE), this indicator can predict the size of the relative error. The relative error refers to the percentage between the actual value and the absolute error. The specific value is , which represents a good model, and , which represents a poor model. The formula is as follows:

[0166]

[0167] Root mean square error, this indicator is the square root of MSE. The calculation formula is as follows:

[0168]

[0169] Compared with the conventional prediction method before the improved model:

[0170] Models MAPE RMSE RNN 6.32 3.12 CNN-GRU 5.27 2.89

[0171] Aiming at the problem that the characteristics of load data are not obvious and the regularity is unstable, a learning model method based on neural network is proposed. The powerful nonlinear mapping and data mining capabilities of the neural network model can effectively solve the problem of inaccurate load forecasting. The present invention constructs the model from four aspects: theoretical analysis, data processing, neural network algorithm and optimization algorithm. On the one hand, theoretical analysis provides strong support for the construction of experimental model. On the other hand, data processing, neural network algorithm and optimization algorithm can effectively construct models for different data sets, which will help to further analyze the research on load forecasting and conduct model evaluation more comprehensively and accurately. In terms of model construction method, the present invention combines CNN and GRU to establish a prediction model based on CNN-GRU. The hybrid model combines the advantages of CNN and GRU models, uses CNN to extract the features of input data, and the GRU network can extract the time dependency in the input data. The grid search method is used to find the optimal parameters to obtain the best model. Combined with the existing power load data, a case analysis is carried out, and multiple different models are established for comparative experiments. The experimental results show that the CNN-GRU model effectively improves the prediction effect of the model.

[0172] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0173] 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 combining similar day factors, characterized in that: include: Step S1, normalizing the meteorological characteristics, constructing an initial data set based on the processed meteorological characteristics, holiday characteristics and load data, performing clustering and association analysis on the initial data set, and obtaining a similar day data set; Step S2, construct a CNN-GRU model, train it with similar day data sets, and adjust the model parameters by grid search method; Step S3: stop model training when the validation set parameter search is completed, evaluate the trained model, search for the next set of parameters, and compare the model effects under various parameters to select the optimal model. Step S4: forecast short-term power load based on the optimal model.

2. A short-term power load forecasting method combining similar day factors according to claim 1, characterized in that: The meteorological data include average wind speed x1, accumulated precipitation x2, temperature x3, accumulated sunshine x4 and average humidity x5, and the load data Y is the maximum load value of the day.

3. A short-term power load forecasting method combining similar day factors according to claim 1, characterized in that: In step S1, the meteorological data is normalized as follows: Based on the MinMaxScaler function, the meteorological data is normalized so that the meteorological data is located in [0,1], specifically: Among them, x i For meteorological data.

4. A short-term power load forecasting method combining similar day factors according to claim 1, characterized in that: In step S1, the initial data is clustered, specifically: Step S11, randomly selecting a sample from the initial data as the initial cluster center; Step S12: Calculate the Euclidean distance between each sample in the initial data and the initial cluster center, and calculate the probability θ(x) of each sample being selected as the next cluster center according to the following formula: Among them, x is the sample, is the sum of the squares of all individual distance measures in set X; Select the next cluster center based on the roulette wheel method; Step S13, repeat step S12 until K cluster centers are selected; Step S14: for each sample in the initial data, calculate the distance from each sample to the K cluster centers, and divide each sample into the class corresponding to the cluster center with the smallest distance; Step S15: Calculate the cluster center C of each class i Step S16: Repeat steps S14-S15 until the location of the cluster center remains unchanged and a rough set is obtained.

5. A short-term power load forecasting method combining similar day factors according to claim 4, characterized in that: The association analysis described in step S1 is performed to obtain a similar day data set, specifically: Step S16: Confirm the factors affecting the power load, the total sequence of factors affecting the i-th day X i for: X i =(x 1i ,x 2i ,...,x ni ) Among them, n is the characteristic value of similar day factor, i=1,2,…,m, m is the number of days of rough set records, and the characteristic subsequence X0 of the day to be predicted is selected as: X0=(x 10 ,x 20 ,...,x n0 ) Step S17: Divide each element in the total sequence and the subsequence by the average value in the respective vectors for averaging. Let the vector after averaging be X′ i and X′0, X′ i The relationship coefficient with X′0 is ξ i(k) : Where k = 1, 2, …, n, |X′ 0(k) -X′ i(k) | is the absolute difference between the total sequence and the subsequence influencing factors after averaging, and ρ is the resolution coefficient; Step S18: Each factor corresponds to a correlation coefficient, so there are n correlation coefficients, and their average value is the correlation degree r i : The data with correlation greater than the preset value are selected to form a data set similar to the day to be predicted.

6. A short-term power load forecasting method combining similar day factors according to claim 1, characterized in that: The training process of the CNN-GRU model in step S2 is: Initialize the CNN-GRU model, calculate through forward propagation, get the initial prediction result, calculate the error between the predicted value and the true load value, and judge whether it converges. If it converges, end the training. If not, backpropagate the error to the model and update the weights of the CNN and GRU models.

7. A short-term power load forecasting method combining similar day factors according to claim 6, characterized in that: The weight update process of the CNN model is: Use k one-dimensional convolution kernels to convolve the multivariate time series x1,x2,...,x T ∈R k The time series of each variable in Extract sequence features to obtain the feature vector of the i-th sequence The calculation process is shown in the formula: represents the jth component of the cth convolution kernel, b is the bias parameter, and f is the activation transformation function.

8. A short-term power load forecasting method combining similar day factors according to claim 6, characterized in that: The weight update process of the GRU model is: Among them, u and r represent the update gate and the reset gate.

9. A short-term power load forecasting method combining similar day factors according to claim 1, characterized in that: The grid search method in step S2 adjusts the model parameters specifically as follows: Build a dictionary of parameters to be adjusted, which contains all the parameters that need to be adjusted in the model and their selectable values; define a classifier, put the parameters that need to be adjusted into the classifier, and get an object clf; build a GridSearchCV object; train the GridSearchCV object and find the optimal parameters.

10. A short-term power load forecasting method combining similar day factors according to claim 1, characterized in that: The specific process of model evaluation is as follows: Mean absolute percentage error: Root Mean Square Error:

Images