Short-term power load prediction method and system considering multi-dimensional feature extraction

Abstract

The invention provides a short-term power load prediction method and system considering multi-dimensional feature extraction, and the method comprises the steps: obtaining regional loads and historical data of related weather of regional influence loads, and dividing a data set; extracting multi-time-scale time sequence features, including trend features, week features and annual features, in the regional load sequence, and constructing an internal feature set; key dependent meteorology with load correlation higher than a set value in the historical meteorological data is searched by using the maximum mutual information coefficient, and an external feature set is constructed; constructing a prediction model fused with a self-attention mechanism, taking the internal feature set and the external feature set as the input of the model, analyzing a prediction value and an actual value to obtain a conditional probability density function, and calculating load condition prediction error probability distribution by taking the load prediction value of the last part of data as the input of the conditional probability density function, and determining a load prediction interval. The prediction result is high in precision.

Description

Technical Field

[0001] The present invention belongs to the field of power load forecasting, and in particular relates to a short-term power load forecasting method and system considering multi-dimensional feature extraction. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Due to the continuous access of a large number of distributed photovoltaics, the continuous promotion of demand response and the improvement of the degree of electricity marketization, the time-varying and volatility of power grid loads have significantly increased. The complex and changeable influencing factors make it more difficult to explore the load rules, resulting in increasing difficulty in its prediction. At the same time, with the increasing pressure on power grid regulation and the rapid advancement of the construction of the power spot market, power grid dispatching agencies attach great importance to short-term load forecasting and have put forward higher requirements for the accuracy, objectivity and automation of forecasting. Therefore, accurate and reliable short-term load forecasting can effectively reduce the uncertainty of load and play an important role in the safe dispatch and operation control of the power system.

[0004] At present, the commonly used load forecasting methods include statistical methods and artificial intelligence methods. Load forecasting based on statistical methods such as linear regression model and integrated moving average autoregressive model is based on the decomposition of time series, and statistical analysis and processing are performed on the decomposed series to achieve the purpose of prediction, but it cannot fully exploit nonlinear characteristics. Artificial intelligence algorithms such as back propagation (BP) neural network and extreme learning machine have simple structures and strong expression capabilities, but cannot fully learn external features such as climate and date. However, deep learning algorithms such as long short-term memory (LSTM) neural network and convolutional neural network can better exploit the internal dependencies of time series and have broad prospects in the field of load forecasting.

[0005] At the same time, existing research has analyzed and summarized the load timing characteristics and dependencies to a certain extent. However, the growth of meteorological sensitive loads and distributed photovoltaics has caused load fluctuations to present new forms, and it is increasingly difficult to accurately grasp the characteristics of load fluctuations. On the one hand, the current exploration of load dependencies is still insufficient, and there are relatively few studies on the correlation between timing characteristics and dependencies; on the other hand, the study of the impact of distributed photovoltaic power generation on load fluctuations is still in its infancy. Due to the difficulty in data acquisition, the analysis results that separate the laws of distributed photovoltaic power generation and load fluctuations are difficult to directly apply in power grid load forecasting. Summary of the invention

[0006] In order to solve the above problems, the present invention proposes a short-term power load forecasting method and system considering multi-dimensional feature extraction. The present invention can solve the problems of low accuracy and incomplete feature extraction of short-term load forecasting technology in the prior art.

[0007] According to some embodiments, the present invention adopts the following technical solutions:

[0008] A short-term power load forecasting method considering multi-dimensional feature extraction includes the following steps:

[0009] Obtain historical data on regional load and relevant meteorological data affecting regional load, and divide the data set into three parts;

[0010] Extract multi-time-scale time series features from regional load series, including trend features, weekly features, and annual features, and construct intrinsic feature sets;

[0011] The maximum mutual information coefficient is used to find the key dependent meteorological data with a load correlation higher than the set value in the historical meteorological data, and an external feature set is constructed;

[0012] Constructing a prediction model integrating a self-attention mechanism, taking the intrinsic feature set and the extrinsic feature set as the input of the model, taking the load prediction value as the output, and using a part of the data set to train the model;

[0013] The trained model is used to process another part of the data to obtain the predicted value. The predicted value and the actual value are analyzed to obtain the conditional probability density function. The load prediction value of the last part of the data is used as the input of the conditional probability density function to calculate the probability distribution of the load conditional prediction error, and then the load prediction interval is determined.

[0014] As an optional implementation method, the specific process of obtaining historical meteorological data related to regional load and regional impact load includes extracting the load data of the target area according to the set frequency and arranging it in chronological order; the obtained historical meteorological data related to the regional impact load includes radiation intensity, temperature, wind speed and wind direction.

[0015] As an optional implementation method, the specific process of dividing the data set into three parts includes using one part of the data as a training set for the prediction model, one part of the data as a statistical source of prediction error distribution of the prediction model, and another part of the data for prediction error probability distribution. The proportions of the three parts of data are 3:4:3.

[0016] As an optional implementation method, the specific process of extracting multi-time scale time series characteristics from the regional load sequence includes: using Prophet to fit the change law of historical load time series data, and decomposing the original time series into trend, cycle, holiday and error components.

[0017] As an optional implementation method, the specific process of using the maximum mutual information coefficient to find the key dependent meteorological factors in the historical meteorological data whose correlation with the load is higher than the set value includes: the maximum mutual information coefficient is used to measure the correlation between the meteorological factors and the load respectively, and the maximum mutual information coefficient is calculated as: Where X and Y are photovoltaic power and meteorological factors; x and y are row and column parameters when gridding; D is D(X,Y), that is, the finite data composed of X and Y; M(D) x,y is the maximum mutual information value of the grid divided by x×y after normalization; B(n) is the upper limit of the number of grids divided; n is the number of samples in the data set D;

[0018] The larger the maximum mutual information coefficient is, the higher the correlation between the two variables is. The meteorological factors that meet the quantity requirements and have the highest correlation with the load and distributed photovoltaic power generation are selected to construct the external feature set.

[0019] As an optional implementation, the prediction model is a long short-term memory network model, including a plurality of cell units, wherein x t represents the input of the cell at time t, c t represents the state of the cell at time t, s t Represents the output of the cell unit; each cell unit contains three control gates, namely the forget gate f, the input gate i and the output o; the forget gate determines the cell unit state c at the previous moment t-1 How much can be saved to c t The input gate determines the current input x t How much can be saved to c t Middle; the output gate determines c t How much can be transferred to s t middle.

[0020] As an optional implementation, the implementation process of the model fusion self-attention mechanism includes: as a self-attention layer, it is used to establish the correlation between time series features. By linking the features between different time steps, the mechanism gives higher weights to the strongly correlated features in the sequence, while weakening the secondary feature information or noise information. The self-attention process is:

[0021]

[0022] In the formula, d is the sample [x 1 ,x 2 ,...,x m ] is the length of each vector obtained by adaptive linear mapping; Q, K, V are the matrices obtained by combining the vectors in parallel.

[0023] As an optional implementation method, the predicted value and the actual value are analyzed to obtain a conditional probability density function, and the load prediction value of the last part of the data is used as the input of the conditional probability density function to calculate the probability distribution of the load conditional prediction error, and then determine the specific process of the load prediction interval including:

[0024] The actual value and the predicted value are counted respectively to obtain the marginal distribution F(x) of the actual load value and the marginal distribution G(y) of the predicted value;

[0025] Calculate the kernel distribution estimates of R and P, and obtain the distribution functions of the actual and predicted load values;

[0026] Use the copulafit function to fit the three Copula function families: normal Copula, Clayton-Copula, and t-Copula. Use the Akaike Information Criterion to evaluate the fitting effects of the three. Select the Copula function with the best goodness of fit based on the minimum AIC value.

[0027] Another part of the data is input into the forecasting model, and the obtained load forecast value is input into the Copula joint probability density function, and the conditional forecast error probability distribution under the known forecast conditions is calculated;

[0028] The estimated interval is obtained by adopting the principle of the narrowest estimated interval, and the estimated interval is converted into the load estimated interval to obtain the load forecast interval.

[0029] As an optional implementation, the conditional probability density function is:

[0030] h XY (e|y)=C XY [F(e+y),G(y)]f(e+y);

[0031] In the formula, x is the actual load value, y is the load forecast value, and h XY (x,y) is the joint probability density function of x and y, function C XY It is called the Copula joint probability density function of the random variables {x, y}, F(x) and G(y) are the marginal distribution functions of x and y, f(x) is the marginal probability density function of x, and the parameter θ of the Copula function is estimated by the maximum likelihood method.

[0032] A short-term power load forecasting system considering multi-dimensional feature extraction, comprising:

[0033] The data processing module is configured to obtain historical data of regional load and relevant meteorological data of regional load influencing, and divide the data set into three parts;

[0034] The intrinsic feature extraction module is configured to extract multi-time scale time series features in the regional load sequence, including trend features, weekly features, and annual features, and construct an intrinsic feature set;

[0035] The external feature extraction module is configured to use the maximum mutual information coefficient to find the key dependent meteorological data with a load correlation higher than a set value in the historical meteorological data, and construct an external feature set;

[0036] A model building and training module is configured to build a prediction model integrating a self-attention mechanism, take the intrinsic feature set and the extrinsic feature set as the input of the model, take the load prediction value as the output, and use a part of the data set to train the model;

[0037] The prediction module is configured to use the trained model to process another part of the data to obtain the predicted value, analyze the predicted value and the actual value to obtain the conditional probability density function, use the load prediction value of the last part of the data as the input of the conditional probability density function, calculate the probability distribution of the load conditional prediction error, and then determine the load prediction interval.

[0038] A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the steps in the above method are completed.

[0039] An electronic device comprises a memory and a processor, and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps in the above method are completed.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] The present invention comprehensively considers the multi-time scale time series characteristics of the load, the influence of key dependent meteorological factors, and the correlation between the actual value and the predicted value of the load, and the prediction result has high accuracy.

[0042] The prediction model of the present invention adds a self-attention mechanism to give higher weights to strongly correlated features in the sequence, highlighting important feature information while weakening secondary feature information or noise information, thereby improving the prediction accuracy of the model.

[0043] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0045] Figure 1It is a flow chart of a method for short-term power load condition prediction error probability distribution considering multi-dimensional feature extraction provided by the first embodiment of the present invention;

[0046] Figure 2 It is a schematic diagram of data set division in a method for short-term power load condition prediction error probability distribution considering multi-dimensional feature extraction provided in the first embodiment of the present invention;

[0047] Figure 3 It is the intrinsic feature extracted by Prophet in a short-term power load condition forecasting error probability distribution method considering multi-dimensional feature extraction provided in the first embodiment of the present invention;

[0048] Figure 4 It is a basic framework diagram of Prophet in a short-term power load condition prediction error probability distribution method considering multi-dimensional feature extraction provided in the first embodiment of the present invention;

[0049] Figure 5 It is a heat map of MIC values ​​between meteorology and photovoltaic power in a short-term power load condition prediction error probability distribution method considering multi-dimensional feature extraction provided in the first embodiment of the present invention;

[0050] Figure 6 It is a network structure diagram of an LSTM model fused with a self-attention mechanism preset in a method for predicting error probability distribution of short-term power load conditions considering multi-dimensional feature extraction provided in the first embodiment of the present invention;

[0051] Figure 7 It is a framework diagram of a multi-dimensional feature LSTM-Copula combined probability prediction model in a short-term power load condition prediction error probability distribution method considering multi-dimensional feature extraction provided in the first embodiment of the present invention;

[0052] Figure 8 It is a conditional prediction error probability distribution step diagram in a short-term power load conditional prediction error probability distribution method considering multi-dimensional feature extraction provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0053] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0054] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0055] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0056] In the absence of conflict, the embodiments in this application and the features in the embodiments may be combined with each other.

[0057] Embodiment 1

[0058] A short-term power load condition prediction method considering multi-dimensional feature extraction, such as Figure 1 As shown, the following steps are included:

[0059] Obtain historical data on regional load and relevant meteorological data affecting regional load, and divide the data set;

[0060] Prophet is used to extract multi-time scale time series features in regional load series, including trend features, weekly features and annual features, and to construct an intrinsic feature set;

[0061] The maximum mutual information coefficient is used to find the key dependent meteorological data with high correlation with load in historical meteorological data, and the external feature set is constructed;

[0062] Construct an LSTM model integrating the self-attention mechanism. Take the above-obtained feature set as input and train it based on the original regional load data to obtain the final output;

[0063] The Copula function is used to explore the probability distribution of load condition prediction errors.

[0064] A detailed introduction is given below.

[0065] First, obtain the historical meteorological data related to the regional load and the regional influencing load, and divide the data set, wherein the acquired load historical data is the load data of the experimental area extracted at a set frequency (in this embodiment, 1 time / 15 minutes), and arranged in chronological order. The acquired relevant meteorological historical data on the regional influencing load include indicators such as irradiation intensity, temperature, wind speed, wind direction, and cloud cover. The data set is divided into three parts: Data set 1 is the training set of the LSTM model, Data set 2 is used for the prediction error distribution statistics of the LSTM model, and Data set 3 is used for Copula probability distribution prediction. The ratio is 3:4:3. Figure 2 shown.

[0066] Secondly, Prophet is used to extract multi-time scale time series features in regional load series, including trend features, weekly features, and annual features, and to construct an intrinsic feature set, such as Figure 3 As shown in the figure, Prophet is a time series analysis tool developed by Taylor et al. that integrates time series decomposition and prediction functions. By fitting the changing rules of historical time series data, the original time series is decomposed into trend, cycle, holiday and error components. The basic framework is as follows Figure 4 Prophet can achieve fast iterative optimization, extract various periodic and holiday effects in time series, provide more regular feature samples for the input of load forecasting model, reduce data complexity, and improve the interpretability of load power consumption characteristics.

[0067] For any regional load sequence L(t), Prophet can decompose it into a combination of multiple regular terms and error terms, as follows:

[0068] L(t)=g(t)+s(t)+h(t)+ε t

[0069] Where: g(t) is the trend term, which is used to fit the non-periodic trend changes in the load sequence; s(t) is the periodic term, which characterizes the periodic changes in the load sequence (such as year, month, week, and day periodicity); h(t) is the holiday term, which reflects the impact of irregular holiday effects; ε t is the error term, which represents the specific variation that the model cannot accommodate and is generally assumed to follow a Gaussian distribution.

[0070] The trend term g(t) includes the saturated growth model and the linear growth model based on the logistic regression function. Here, the segmented model based on linear regression is used to fit the trend change law of regional load, as follows:

[0071] g(t)=(k+α T (t)δ)t+(M+α T (t)γ)

[0072]

[0073] Where: k is the basic growth rate of regional load; δ is the load growth rate adjustment vector; α(t) is the indicator function; θ j (t) is the value of α(t) at the jth mutation point; s j is the time when the jth mutation point in the regional load sequence is located; M is the offset parameter; γ is the smoothing offset at the mutation point; the growth rate of the regional load at time t is k+α T (t)δ.

[0074] The periodicity s(t) uses Fourier series to simulate the periodicity of the time series as follows:

[0075]

[0076] β={a 1 ,b 1 ,a 2 ,b 2 ,...,a N ,b N}

[0077] Where: P is the period, in days; β is the smoothing coefficient a γ and b γ The collection satisfies the normal distribution; N is the smoothing coefficient a γ or b γ ; γ is the serial number of the smoothing coefficient.

[0078] For the holiday term h(t), Prophet assumes that every day during the holiday is independent of each other, as follows:

[0079] h(t)=Z(t)κ

[0080] Z(t)=[1(t∈D 1 ),1(t∈D 2 ),...,1(t∈D L )]

[0081] κ~N(0,v 2 )

[0082] Where: Z(t) is the regression matrix; D l is the set of past and future dates of the first holiday, l = 1, 2, ..., L, with a total of L days; κ is the prior variation parameter corresponding to the holiday, which satisfies the normal distribution; v is the influence of the holiday, and the larger v is, the greater the influence of the holiday.

[0083] Step 3: Use the maximum mutual information coefficient to find the key dependent meteorological data with high correlation with load in the historical meteorological data, and construct the external feature set. The maximum mutual information coefficient (MIC) was proposed by Reshef et al. in 2011 and is developed on the basis of mutual information (MI). Mutual information represents a useful information measure in information theory, which indicates the amount of information contained in a random variable about another random variable. It can also be understood as the uncertainty of a random variable reduced by another known random variable. The mutual information calculation formula is:

[0084]

[0085] Where: I(X,Y) is the mutual information of variables X and Y; P(x,y) is the joint density function of X and Y; P(x) and P(y) are the marginal density functions of X and Y, respectively. This embodiment takes the 2020 regional load data as an example, where X and Y are the regional load throughout the year and the meteorological data of the corresponding data period throughout the year.

[0086] The scatter plot composed of the finite data set D(X,Y) is gridded according to the X-axis and Y-axis with x rows and y columns to obtain an x×y grid G, and D| G It represents the distribution of data set D on grid G ​​(i.e., probability density), which can be obtained by the proportion of points falling into grid G ​​in data set D. Using different methods to divide grid G ​​will result in different D| G Probability distribution. Calculate the mutual information under different distributions and take the maximum value:

[0087] I * (D,x,y)=maxI(D| G )

[0088] Where: I * (D,x,y) is the maximum mutual information value of the grid divided into x×y.

[0089] In order to facilitate comparison and analysis between data of different units or magnitudes, the data are normalized:

[0090]

[0091] Where: M(D) x,y is the normalized I * (D,x,y).

[0092] M(D) obtained by dividing the grid with different x×y values x,y The maximum value is the maximum mutual information coefficient (MIC value) between variables X and Y:

[0093]

[0094] Where: B(n) is the upper limit of the number of grids; n is the number of samples in the data set D. X, Y are photovoltaic power and meteorological factors; x, y are row and column parameters when gridding; D is D(X, Y), that is, the finite data composed of X and Y; M(D) x,y It is the maximum mutual information value of the grid divided into x×y after normalization.

[0095] If B(n) is too high, the results will be conservative, and if B(n) is too low, the ability to retrieve complex relationships will be weakened. According to empirical tests: B(n) = n α, parameter α is determined by the sample size, and this patent takes α = 0.6. The MIC value reflects the degree of association between two variables, and the value range is 0 to 1. When the MIC value tends to 0, it can be considered that the association between the two variables is extremely weak or non-existent, and when the MIC value tends to 1, it can be considered that the association between the two variables is extremely high.

[0096] In some embodiments, the Figure 5 The heat map shown in the figure selects the largest 2 to 3 MIC values ​​to obtain the meteorological factors with the highest correlation with load and distributed photovoltaic power generation, and constructs an external feature set as the input of the subsequent LSTM model.

[0097] Then, an LSTM model integrating the self-attention mechanism is constructed. The feature set obtained in the above two steps is used as input and trained based on the original regional load dataset 1. The long short-term memory neural network (LSTM) is a variant of the recurrent neural network RNN. Due to the vanishing gradient, RNN can only have short-term memory, while LSTM overcomes the vanishing gradient problem by introducing well-defined memory units and gate structures. Therefore, compared with RNN, LSTM has a stronger ability to store and access historical sequence information.

[0098] like Figure 6 As shown, the hidden layer of the LSTM neural network has three inputs, namely the cell state c at the previous moment t-1 , the output value s at the previous moment t-1 and the current input value x t The management and control of the cell state c is the key to the LSTM neural network's ability to memorize information for a long time. LSTM implements this control by introducing three gate structures: the forget gate, the input gate, and the output gate.

[0099] The forget gate is used to control the current cell state c t The cell state c of the previous moment is saved in t-1 The information that needs to be saved is expressed as follows:

[0100] f t =σ(W f ·[s t-1 ,x t ]+b f )

[0101] Where W f is the weight matrix of the forget gate, σ is the sigmoid activation function of the forget gate, b f is the bias of the forget gate, x t is the input at the current moment, s t-1 is the unit state at the previous moment, and the final ft is a value in [0,1]. t =1 means that the unit state at the previous moment is fully remembered, otherwise if f t = 0 means that the unit state at the previous moment is completely forgotten, f t The value of is generally (0,1), and only the information that needs to be saved in the unit state at the previous moment is memorized.

[0102] The input gate controls the current cell state c t How much current input x is saved in t The information is expressed as follows:

[0103] i t =σ(W i ·[s t-1 ,x t ]+b i )

[0104] Where W i is the weight matrix of the input gate, σ is the sigmoid activation function of the input gate, b i is the bias of the input gate, x t is the input at the current moment, s t-1 is the unit state at the previous moment, and the final i t is a value in [0,1]. t =1 means that all input at this time is remembered, otherwise if i t = 0 means that all input at this time is forgotten, i t The value of is generally (0,1), and only the information that needs to be saved in the input at this time is memorized.

[0105] Get the candidate value vector of the current cell state is to get the current cell state c t The premise is The expression is as follows:

[0106]

[0107] Where W c is the weight matrix, tanh is the activation function, b c is the bias, x t is the input at the current moment, s t-1 is the unit state at the previous moment, and the final is a value in the range [0,1].

[0108] Current cell status c t The expression is as follows:

[0109]

[0110] In the formula, c t-1 is the unit state at the previous moment, i t 、f t is the value at the current moment, and can be obtained through the above formula, which can integrate the current memory and long-term memory to obtain the current unit state c t .

[0111] The output gate controls the current output s t The current cell state c is saved in t The amount of information in is expressed as follows:

[0112] o t =σ(W o ·[s t-1 ,x t ]+b o )

[0113] s t =o t *tanh(c t )

[0114] Where W o is the weight matrix of the output gate, σ is the sigmoid activation function of the output gate, b o is the bias of the output gate, x t is the input at the current moment, s t-1 is the unit state at the previous moment, and the obtained o t is a value in the range [0,1].

[0115] Finally, the current unit is combined with the output gate to obtain o t Integrate to get the current output s t .

[0116] The self-attention mechanism is a special attention mechanism, also known as internal attention, which can characterize the correlation between different features within the input sequence and assign weights. The implementation of this mechanism is to use the self-attention layer to establish the correlation between time series features, and to avoid the problem of information forgetting caused by too long time steps by linking features between different time steps. This mechanism gives higher weights to strongly correlated features in the sequence, highlighting important feature information, while weakening secondary feature information or noise information, thereby improving the prediction accuracy of the model.

[0117] First, for m feature sequences [x 1 ,x 2 ,...,x m ] is adaptively linearly mapped and converted into three vectors q of length d i ,k i ,vi , and then combine the three vectors in parallel into matrices Q, K, and V. Calculate the matrix S according to the following formula. S is the weight set between m sequences, and then normalize it to become S n [S n1 ,S n2 ,...,S nm ] T .

[0118] S=QK T

[0119]

[0120] According to the following formula, the weight is converted into probability form using the Soft max function P = [P 1 ,P 2 ,...,P m ].

[0121]

[0122] The weight matrix Z is obtained according to the following formula:

[0123] Z=V·P

[0124] Then, the self-attention process is as follows:

[0125]

[0126] Where: d is the sample [x 1 ,x 2 ,...,x m ] are the lengths of the three vectors obtained by adaptive linear mapping; Q, K, V are the matrices obtained by combining the three vectors in parallel.

[0127] In summary, the input of the LSTM model integrating the self-attention mechanism is: the trend characteristics, weekly characteristics and annual characteristic data of the regional load series extracted by Prophet, and the numerical weather forecast data of the key dependent meteorological conditions screened by MIC. The original load data set 1 is used to train the LSTM model to provide effective prediction values ​​for the Copula probability distribution prediction. Its advantages are that it comprehensively considers the multi-time scale time series characteristics of the load, the influence of key dependent meteorological factors, and the correlation between the actual load value and the predicted value, and the prediction result has high accuracy.

[0128] Finally, the Copula function is used to explore the probability distribution of load condition prediction error. The Copula function was first proposed by Sklar in 1959 and can describe the correlation between multivariate random variables. Here, the binary Copula function is used to analyze the dependence between the actual load value and the predicted value, avoiding the drawback of the lack of sufficient constraints in the traditional normal distribution method. For the marginal probability density functions f(x) and g(y) of the random variables {x, y}, their joint probability density function h XY (x,y) has a unique Copula function that connects them, and the expression is as follows:

[0129] h XY (x,y)=C XY [F(x),G(y)]f(x)g(y)

[0130] In the formula, function C XY It is called the Copula joint probability density function of the random variables {x, y}. The main parameter θ of the Copula function can be estimated by the maximum likelihood method.

[0131] The load forecast error is

[0132] e=xy

[0133] Where x is the actual load value and y is the predicted load value.

[0134] Therefore, the conditional probability density function of the prediction error can be expressed as:

[0135] h XY (e|y)=C XY [F(e+y),G(y)]f(e+y)

[0136] From the above formula, we can know that the load condition prediction error can be calculated by calculating the Copula joint probability density of the actual load value and the predicted value and the probability density of the actual value.

[0137] The steps for conditional forecast error probability distribution are as follows:

[0138] like Figure 7 , Figure 8 As shown in the figure, the actual load value of data set 2 is recorded as R, and the load prediction value obtained by inputting the LSTM model into data set 2 is recorded as P. The actual value and the predicted value are statistically analyzed to obtain the marginal distribution F(x) of the actual load value and the marginal distribution G(y) of the predicted value.

[0139] Calculate the kernel distribution estimates of R and P, and obtain the distribution functions of the actual and predicted load values;

[0140] The copulafit function was used to fit the three Copula function families: normal Copula, Clayton-Copula, and t-Copula. The Akaike Information Criterion (AIC) was used to evaluate the fitting effects of the three. The Copula function with the best fitting goodness of fit was selected based on the minimum AIC value.

[0141] The load forecast value obtained by inputting data set 3 into the LSTM model is input into the Copula joint probability density function, and the conditional forecast error probability distribution under the known forecast conditions is calculated;

[0142] The estimated interval [η 1 ,η 2 ]:

[0143]

[0144] Where: 1-α is the modeling confidence.

[0145] The estimated interval is then converted into a load estimation interval to obtain the load forecast interval.

[0146] Embodiment 2

[0147] A short-term power load forecasting system considering multi-dimensional feature extraction, comprising:

[0148] The data processing module is configured to obtain historical data of regional load and relevant meteorological data of regional load influencing, and divide the data set into three parts;

[0149] The intrinsic feature extraction module is configured to extract multi-time scale time series features in the regional load sequence, including trend features, weekly features, and annual features, and construct an intrinsic feature set;

[0150] The external feature extraction module is configured to use the maximum mutual information coefficient to find the key dependent meteorological data with a load correlation higher than a set value in the historical meteorological data, and construct an external feature set;

[0151] A model building and training module is configured to build a prediction model integrating a self-attention mechanism, take the intrinsic feature set and the extrinsic feature set as the input of the model, take the load prediction value as the output, and use a part of the data set to train the model;

[0152] The prediction module is configured to use the trained model to process another part of the data to obtain the predicted value, analyze the predicted value and the actual value to obtain the conditional probability density function, use the load prediction value of the last part of the data as the input of the conditional probability density function, calculate the probability distribution of the load conditional prediction error, and then determine the load prediction interval.

[0153] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0154] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0155] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0156] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0157] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principle of the present invention without creative labor shall be included in the protection scope of the present invention.

Claims

1. A short-term power load forecasting method considering multi-dimensional feature extraction, characterized in that: The following steps are involved: Obtain historical data on regional load and relevant meteorological data affecting regional load, and divide the data set into three parts; Extract multi-time-scale time series features from regional load series, including trend features, weekly features, and annual features, and construct intrinsic feature sets; The maximum mutual information coefficient is used to find the key dependent meteorological data with a load correlation higher than the set value in the historical meteorological data, and an external feature set is constructed; Constructing a prediction model integrating a self-attention mechanism, taking the intrinsic feature set and the extrinsic feature set as the input of the model, taking the load prediction value as the output, and using a part of the data set to train the model; The trained model is used to process another part of the data to obtain the predicted value. The predicted value and the actual value are analyzed to obtain the conditional probability density function. The load prediction value of the last part of the data is used as the input of the conditional probability density function to calculate the probability distribution of the load conditional prediction error, and then determine the load prediction interval.

2. A short-term power load forecasting method considering multi-dimensional feature extraction as claimed in claim 1, characterized in that: The specific process of obtaining historical meteorological data related to regional load and regional influencing load includes extracting load data of the target area according to the set frequency and arranging them in chronological order; the obtained historical meteorological data related to regional influencing load includes radiation intensity, temperature, wind speed and wind direction.

3. A short-term power load forecasting method considering multi-dimensional feature extraction as claimed in claim 1, characterized in that: The specific process of dividing the data set into three parts includes using one part of the data as the training set of the prediction model, one part of the data as the statistical source of the prediction error distribution of the prediction model, and another part of the data for the prediction error probability distribution. The proportions of the three parts of data are 3:4:

3.

4. A short-term power load forecasting method considering multi-dimensional feature extraction as claimed in claim 1, characterized in that: The specific process of extracting multi-time-scale time series characteristics from regional load sequences includes: using Prophet to fit the changing patterns of historical load time series data and decomposing the original time series into trend, cycle, holiday and error components.

5. A short-term power load forecasting method considering multi-dimensional feature extraction as claimed in claim 1, characterized in that: The specific process of using the maximum mutual information coefficient to find the key dependent meteorological factors in the historical meteorological data whose correlation with the load is higher than the set value includes: the maximum mutual information coefficient is used to measure the correlation between meteorological factors and load respectively, and the maximum mutual information coefficient is calculated as: Where X and Y are photovoltaic power and meteorological factors; x and y are row and column parameters when gridding; D is D(X,Y), that is, the finite data composed of X and Y; M(D) x,y is the maximum mutual information value of the grid divided by x×y after normalization; B(n) is the upper limit of the number of grids divided; n is the number of samples in the data set D; The larger the maximum mutual information coefficient is, the higher the correlation between the two variables is. The meteorological factors that meet the quantity requirements and have the highest correlation with the load and distributed photovoltaic power generation are selected to construct the external feature set.

6. A short-term power load forecasting method considering multi-dimensional feature extraction as claimed in claim 1, characterized in that: The prediction model is a long short-term memory network model, including multiple cell units, where x t represents the input of the cell at time t, c t represents the state of the cell at time t, s t Represents the output of the cell unit; each cell unit contains three control gates, namely the forget gate f, the input gate i and the output o; the forget gate determines the cell unit state c at the previous moment t-1 How much can be saved to c t The input gate determines the current input x t How much can be saved to c t Middle; the output gate determines c t How much can be transferred to s t middle.

7. A short-term power load forecasting method considering multi-dimensional feature extraction as claimed in claim 1, characterized in that: The implementation process of the model integrating the self-attention mechanism includes: as a self-attention layer, it is used to establish the correlation between time series features. By linking the features between different time steps, the mechanism gives higher weights to the strongly correlated features in the sequence, while weakening the secondary feature information or noise information. The self-attention process is: Where d is the sample [x1,x2,...,x m ] is the length of each vector obtained by adaptive linear mapping; Q, K, V are the matrices obtained by combining the vectors in parallel.

8. A short-term power load forecasting method considering multi-dimensional feature extraction as claimed in claim 1, characterized in that: The predicted value and the actual value are analyzed to obtain the conditional probability density function. The load prediction value of the last part of the data is used as the input of the conditional probability density function to calculate the probability distribution of the load conditional prediction error, and then the specific process of determining the load prediction interval includes: The actual value and the predicted value are counted respectively to obtain the marginal distribution F(x) of the actual load value and the marginal distribution G(y) of the predicted value; Calculate the kernel distribution estimates of R and P, and obtain the distribution functions of the actual and predicted load values; Use the copulafit function to fit the three Copula function families: normal Copula, Clayton-Copula, and t-Copula. Use the Akaike Information Criterion to evaluate the fitting effects of the three. Select the Copula function with the best goodness of fit based on the minimum AIC value. Another part of the data is input into the forecasting model, and the obtained load forecast value is input into the Copula joint probability density function, and the conditional forecast error probability distribution under the known forecast conditions is calculated; The estimated interval is obtained by adopting the principle of the narrowest estimated interval, and the estimated interval is converted into the load estimated interval to obtain the load forecast interval.

9. A short-term power load forecasting method considering multi-dimensional feature extraction as claimed in claim 1 or 8, characterized in that: The conditional probability density function is: h XY (e|y)=C XY [F(e+y),G(y)]f(e+y); In the formula, x is the actual load value, y is the predicted load value, and h XY (x,y) is the joint probability density function of x and y, function C XY It is called the Copula joint probability density function of the random variables {x, y}, F(x) and G(y) are the marginal distribution functions of x and y, f(x) is the marginal probability density function of x, and the parameter θ of the Copula function is estimated by the maximum likelihood method.

10. A short-term power load forecasting system considering multi-dimensional feature extraction, characterized in that: include: The data processing module is configured to obtain historical data of regional load and relevant meteorological data of regional load influencing, and divide the data set into three parts; The intrinsic feature extraction module is configured to extract multi-time scale time series features in the regional load sequence, including trend features, weekly features, and annual features, and construct an intrinsic feature set; The external feature extraction module is configured to use the maximum mutual information coefficient to find the key dependent meteorological data with a load correlation higher than a set value in the historical meteorological data, and construct an external feature set; A model building and training module is configured to build a prediction model integrating a self-attention mechanism, take the intrinsic feature set and the extrinsic feature set as the input of the model, take the load prediction value as the output, and use a part of the data set to train the model; The prediction module is configured to use the trained model to process another part of the data to obtain the predicted value, analyze the predicted value and the actual value to obtain the conditional probability density function, use the load prediction value of the last part of the data as the input of the conditional probability density function, calculate the probability distribution of the load conditional prediction error, and then determine the load prediction interval.

Images