Power load medium and short term prediction method based on multi-expert model

Abstract

The invention discloses a power load medium and short term prediction method based on a multi-expert model. In a power system, accurate prediction of medium and short-term power loads is crucial for effective energy management. A traditional power load prediction model is often limited by complex variables and nonlinear relations, so that a more innovative and efficient method is needed. According to the medium and short term power load prediction method based on the multi-expert model provided by the invention, the comprehensive analysis capability of various influence factors is improved by integrating professional knowledge in multiple fields. The power load data shows periodic characteristics of hour level, day level, weekly level and month level, a multi-expert model is used, different experts model different power load characteristics respectively, and the accuracy of power load prediction in a power system can be improved.

Description

Technical Field

[0001] The invention relates to a medium- and short-term power load forecasting method based on a multi-expert model, and belongs to the technical field of power load forecasting. Background Art

[0002] Power system load forecasting is one of the key technologies in power system operation and planning. Its forecasting accuracy directly affects multiple links such as grid dispatching, power generation planning, demand-side management and new energy consumption. Medium- and short-term power load forecasting, especially the power demand forecast for the next few hours to weeks, is crucial to the stability and economy of the power system. With the rapid growth of energy demand, power load forecasting faces increasingly complex challenges, including climate change, fluctuations in socio-economic factors, and large-scale access to new energy. These uncertainties and complexities have caused traditional load forecasting methods to show limitations in dealing with multivariable and nonlinear problems.

[0003] Traditional power load forecasting methods usually include time series analysis, regression analysis, and neural networks. Although time series methods such as the ARIMA model can effectively capture the time-dependent characteristics of power load in some cases, their ability to predict long-term and nonlinear load characteristics is limited. In addition, although simple regression methods can take into account the influence of multiple external factors, they also have certain limitations when dealing with complex multidimensional data sets and nonlinear relationships. Although the neural network method has powerful nonlinear mapping capabilities, a single model is prone to overfitting problems and it is difficult to fully analyze the multi-level and periodic characteristics of load data. Therefore, in practical applications, especially in complex and changeable power systems, traditional methods cannot guarantee high accuracy and reliability of predictions.

[0004] The core idea of ​​the multi-expert model is to introduce "experts" from multiple fields, each of whom models a specific type of problem, and uses certain rules or algorithms to comprehensively evaluate and make decisions based on the output of these experts. This method can effectively overcome the limitations of a single model and improve the accuracy and robustness of the forecast results. Especially in power load forecasting, different experts can model different periodic and trend characteristics, thereby more comprehensively analyzing the inherent laws in complex load data.

[0005] In modern power systems, accurate medium- and short-term load forecasting is crucial for optimizing energy resource allocation and improving grid operation efficiency. The forecasting method based on a multi-expert model can not only effectively integrate the advantages of multiple forecasting models, but also make full use of massive data and information on external influencing factors to provide important support for the safe and economical operation of the power system. Therefore, developing an innovative and efficient medium- and short-term power load forecasting method based on a multi-expert model has important technical significance and application value. Summary of the invention

[0006] The purpose of the present invention is to provide a method for medium- and short-term power load forecasting based on a multi-expert model. The multi-expert model is used to overcome the limitations of a single model. Different "experts" can respectively model different periodic and trend characteristics to improve the accuracy and robustness of power load forecasting.

[0007] The purpose of the present invention is achieved through the following technical solutions:

[0008] A method for short- to medium-term power load forecasting based on a multi-expert model comprises the following steps:

[0009] Step 1: Training data preparation

[0010] The historical power load data is divided into four groups: hourly cycle, daily cycle, weekly cycle, and monthly cycle. For the current prediction point, the data of the previous α hours are used as hourly cycle features, the data of the previous β days are used as daily cycle features, the data of the previous γ weeks are used as weekly cycle features, and the data of the previous δ months are used as monthly cycle features. The hourly cycle features are extracted at hourly intervals, the daily cycle features are extracted at daily intervals, the weekly cycle features are extracted at weekly intervals, and the monthly cycle features are extracted at monthly intervals. The inputs of different cycle data are:

[0011]

[0012] Where, X h is the hourly periodic characteristic, x t is the current prediction point, x t-α is the data of the previous α hours, x t-α+1 is the data of the previous α-1 hour; X d is the daily cycle characteristic, x t-β*24 is the data of the previous β days, x t-(β-1)*24 is the data of the previous β-1 day; X w is the periodic characteristic, x t-γ*24*7 is the data of the previous γ weeks, x t-(γ-1)*24*7 is the data of the previous γ-1 week; X m is the monthly cycle characteristic x t-δ*24*30 is the data of the previous δ months, x t-(δ-1)*24*30 It is the data of the previous δ-1 month;

[0013] Step 2: TCN network construction

[0014] The TCN network consists of multiple residual block networks. Each residual block has two dilated causal convolutions. After each dilated causal convolution, the weight is normalized, and the ReLU nonlinear activation function f(x) = max(0, x) and Dropout regularization are used. At the same time, a residual connection is performed every two layers of dilated causal convolutions.

[0015] Step 3: MLP network construction

[0016] An MLP network consists of two layers. The first layer consists of two linear layers, and the second layer is a linear layer network:

[0017]

[0018] In the formula, It is element-by-element multiplication, SiLU is a nonlinear activation function, Linear is a linear layer function, and the formulas of SiLU and Linear are as follows:

[0019]

[0020] In the formula, W is the weight parameter, b is the bias vector parameter, and the sigmoid formula is as follows:

[0021]

[0022] Step 4: Sigmoid gated network construction

[0023] The Sigmoid network controls the weight of the previous network layer's contribution to the next network layer. The Sigmoid gating branch controls the weight of the TCN network. The Sigmoid gating branch consists of an MLP network and a Sigmoid activation function:

[0024]

[0025] Where G(x) is the gating network, It is element-by-element multiplication, Sigmoid is the activation function, MLP and TCN correspond to MLP network and TCN network respectively;

[0026] Step 5: Construction of multi-expert load forecasting model

[0027] The multi-expert load forecasting model is composed of a TCN network, an MLP network, and a Sigmoid gating network. Four different TCN networks receive the hourly cycle feature X h , daily cycle characteristics X d , Weekly characteristics X w 、Monthly cycle characteristics X mAfter the input, the outputs of the four TCN networks are obtained. After the outputs of the TCN networks pass through their respective Sigmoid gating networks, the gated network outputs corresponding to the hourly periodic features, daily periodic features, weekly periodic features, and monthly periodic features are obtained. The four groups of gated network outputs are spliced ​​together to obtain the gated network output; after the gated network output passes through a linear layer and softmax calculation, the topk expert index is selected according to the softmax result, and the selected MLP network is selected according to the index to participate in the calculation, and the gated network output is passed as input to the selected MLP network, and the weight calculated by the above softmax is accumulated and then accumulated to obtain the output of the multi-expert network;

[0028] During the training data preparation process, the input data was divided into four groups: hourly cycle, daily cycle, weekly cycle, and monthly cycle. Each group of input data passed through a different TCN network and Sigmoid gating network, and the outputs of the four groups of gating networks were as follows:

[0029]

[0030] In the formula, x h is the hourly period input, x d It is daily cycle data, x w is the weekly data, x m is the monthly cycle data, G h , G d , G w , G m is the gating network output corresponding to the periodic data;

[0031] After the gating network output results of the four sets of periodic features are spliced ​​together, they pass through different MLP expert networks. Each MLP expert network models different input features, and only the topk experts are selected to participate in the network calculation to obtain the final prediction result:

[0032] G = concat(G h ,G d ,G w ,G m )

[0033] weight,index=topk(softmax(WG+b))

[0034] In the formula, concat is the matrix concatenation calculation, W is the weight parameter of the linear layer, b is the bias vector of the linear layer, topk is the largest k numbers in the softmax calculation result, the largest k numerical results are represented by the weight vector, and the corresponding index value is represented by the index vector. The value of index is the selected expert index. The softmax calculation formula is as follows:

[0035]

[0036] In the formula, x is a set of vectors, x i is the i-th value in the vector;

[0037] After calculating the selected expert index and the corresponding weight, use the selected expert MLP network to perform calculations, and sum the results of each expert network to get the final result:

[0038]

[0039] In the formula, y is the output of the multi-expert model, index is the index result of topk, weight is the weight value corresponding to topk, G is the output of the gating network of the previous layer, and MLP i is the i-th expert network;

[0040] Step 6: Multi-expert model training and evaluation

[0041] The Pytorch framework is used for model training and prediction, and the optimal model is trained through the gradient descent method.

[0042] The purpose of the present invention can also be further achieved by the following technical measures:

[0043] In the aforementioned medium- and short-term power load forecasting method based on a multi-expert model, the TCN network is replaced by an LSTM or CNN network.

[0044] In the above-mentioned medium- and short-term power load forecasting method based on a multi-expert model, the error function for multi-expert model training and evaluation in step 6 uses the root mean square error RMSE and the mean absolute error MAE. The error function formula is as follows:

[0045]

[0046] Where N is the number of samples, y i is the actual load value, is the predicted load value of the model.

[0047] Compared with the prior art, the beneficial effects of the present invention are: the use of a multi-expert model can increase the robustness of the model, different experts are responsible for modeling different periodic characteristics, and through dynamic expert selection, the prediction results are made more accurate and reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is the TCN model structure diagram;

[0049] Figure 2 It is the structure diagram of the MLP model;

[0050] Figure 3 It is the structure diagram of Sigmoid gated network;

[0051] Figure 4 It is a diagram of a multi-expert model for power load forecasting. DETAILED DESCRIPTION

[0052] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0053] The method for medium- and short-term power load forecasting based on a multi-expert model provided by the present invention comprises the following steps:

[0054] Step 1: Training data preparation

[0055] The historical power load data is divided into four groups: hourly cycle, daily cycle, weekly cycle, and monthly cycle, and input into the multi-expert model network respectively. For the current prediction point, the data of the previous α hours of the point are used as hourly cycle features, the data of the previous β days are used as daily cycle features, the data of the previous γ weeks are used as weekly cycle features, and the data of the previous δ months are used as monthly cycle features. Hourly cycle features are extracted at hourly intervals, daily cycle features are extracted at daily intervals, weekly cycle features are extracted at weekly intervals, and monthly cycle features are extracted at monthly intervals. The input of different cycle data are:

[0056]

[0057] Where, X h is the hourly periodic characteristic, x t is the current prediction point, x t-α is the data of the previous α hours, x t-α+1 is the data of the previous α-1 hour, and so on. Xd is the daily cycle feature, x t-β*24 is the data of the previous β days, x t-(β-1)*24 is the data of the previous β-1 day, and so on. w is the periodic characteristic, x t-γ*24*7 is the data of the previous γ weeks, x t-(γ-1)*24*7 is the data of the previous γ-1 week, and so on. m is the monthly cycle characteristic x t-δ*24*30 is the data of the previous δ months, x t-(δ-1)*24*30 It is the data of the previous δ-1 month, and so on.

[0058] Step 2: TCN network construction

[0059] The TCN network structure is as follows Figure 1As shown in the figure, the TCN network consists of multiple residual block networks. Each residual block has two dilated causal convolutions. After each dilated causal convolution, the weights are normalized, and the ReLU nonlinear activation function f(x) = max(0, x) and Dropout regularization are used. At the same time, a residual connection is performed every two layers of dilated causal convolution.

[0060] Step 3: MLP network construction

[0061] The MLP network structure is as follows Figure 2 As shown. An MLP network is mainly composed of two layers of networks. The first layer consists of two linear layers, and the second layer is a linear layer network:

[0062]

[0063] In the formula, It is element-by-element multiplication, SiLU is a nonlinear activation function, Linear is a linear layer function, and the formulas of SiLU and Linear are as follows:

[0064]

[0065] In the formula, W is the weight parameter, b is the bias vector parameter, and the sigmoid formula is as follows:

[0066]

[0067] Step 4: Sigmoid gated network construction

[0068] Sigmoid gating network Figure 3 As shown in the figure. The Sigmoid network mainly controls the weight of the previous network layer to the next network layer. Based on the TCN network model, a Sigmoid gating branch is added to control the weight of the TCN network. The Sigmoid gating branch consists of an MLP network and a Sigmoid activation function.

[0069]

[0070] Where G(x) is the gating network, It is element-by-element multiplication, Sigmoid is the activation function, MLP and TCN correspond to MLP network and TCN network respectively.

[0071] The function of the gating unit is to control the contribution of different input data. The value range of Sigmoid output is (0,1). The closer Sigmoid is to 1, the greater the contribution of the input feature. The closer Sigmoid is to 0, the smaller the contribution of the input feature.

[0072] Step 5: Construction of multi-expert load forecasting model

[0073] The multi-expert load forecasting model is composed of the above-mentioned TCN network, MLP network, and Sigmoid gating network. Figure 4 As shown. The hourly period feature X is received by 4 different TCN networks respectively. h , daily cycle characteristics X d , Weekly characteristics X w 、Monthly cycle characteristics X m After the input, the outputs of the four TCN networks are obtained. After the outputs of the TCN networks pass through their respective Sigmoid gating networks, the gated network outputs corresponding to the hourly periodic features, daily periodic features, weekly periodic features, and monthly periodic features are obtained. The four groups of gated network outputs are spliced ​​together to obtain the gated network output. After the gated network output passes through a linear layer and softmax calculation, the topk expert index is selected according to the softmax result. According to the index, the selected MLP network is selected to participate in the calculation, and the gated network output is passed as input to the selected MLP network, and the weight calculated by the above softmax is accumulated and multiplied to obtain the output of the multi-expert network.

[0074] During the training data preparation process, the input data was divided into four groups: hourly cycle, daily cycle, weekly cycle, and monthly cycle. Each group of input data passed through a different TCN network and Sigmoid gating network to obtain the output of four groups of gating networks.

[0075]

[0076] In the formula, x h is the hourly period input, x d It is daily cycle data, x w is the weekly data, x m is the monthly cycle data, G h , G d , G w , G m is the gating network output corresponding to the periodic data.

[0077] After the gated network output results of the four sets of periodic features are spliced ​​together, they pass through different MLP expert networks. Each MLP expert network models different input features, and only the topk experts are selected to participate in the network calculation to obtain the final prediction results.

[0078] G = concat(G h ,G d ,G w ,G m )

[0079] weight,index=topk(softmax(WG+b))

[0080] In the formula, concat is the matrix concatenation calculation, W is the weight parameter of the linear layer, b is the bias vector of the linear layer, topk is the largest k numbers in the softmax calculation result, the largest k numerical results are represented by the weight vector, and the corresponding index value is represented by the index vector. The value of index is the selected expert index. The softmax calculation formula is as follows:

[0081]

[0082] In the formula, x is a set of vectors, x i is the i-th value in the vector.

[0083] After calculating the selected expert index and the corresponding weight, the selected expert MLP network is used for calculation, and the results of each expert network are summed to obtain the final result.

[0084]

[0085] In the formula, y is the output of the multi-expert model, index is the index result of topk, weight is the weight value corresponding to topk, G is the output of the gating network of the previous layer, and MLP i is the i-th expert network.

[0086] Using a multi-expert model can increase the robustness of the model. Different experts are responsible for modeling different periodic characteristics. Through dynamic expert selection, the prediction results can be more accurate and reliable.

[0087] In the multi-expert model method of this article, the TCN network can be replaced by other time series prediction networks such as LSTM and CNN. This method is applicable to other time series prediction networks.

[0088] Step 6: Multi-expert model training and evaluation

[0089] The Pytorch framework is used for model training and prediction, and the optimal model is trained through the gradient descent method.

[0090] The error function uses the root mean square error (RMSE) and the mean absolute error (MAE). The error function formula is as follows:

[0091]

[0092]

[0093] Where N is the number of samples, y i is the actual load value, is the predicted load value of the model.

[0094] The model is evaluated using two indicators, RMSE and MAE. The smaller the RMSE and MAE values, the higher the prediction accuracy of the model.

[0095] In addition to the above embodiments, the present invention may also have other implementation modes. Any technical solutions formed by equivalent replacement or equivalent transformation shall fall within the protection scope required by the present invention.

Claims

1. A method for medium- and short-term power load forecasting based on a multi-expert model, characterized in that: The following steps are involved: Step 1: Training data preparation The historical power load data is divided into four groups: hourly cycle, daily cycle, weekly cycle, and monthly cycle. For the current prediction point, the data of the previous α hours are used as hourly cycle features, the data of the previous β days are used as daily cycle features, the data of the previous γ weeks are used as weekly cycle features, and the data of the previous δ months are used as monthly cycle features. The hourly cycle features are extracted at hourly intervals, the daily cycle features are extracted at daily intervals, the weekly cycle features are extracted at weekly intervals, and the monthly cycle features are extracted at monthly intervals. The inputs of different cycle data are: Where, X h is the hourly periodic characteristic, x t is the current prediction point, x t-α is the data of the previous α hours, x t-α+1 is the data of the previous α-1 hour; X d is the daily cycle characteristic, x t-β*24 is the data of the previous β days, x t-(β-1)*24 is the data of the previous β-1 day; X w is the periodic characteristic, x t-γ*24*7 is the data of the previous γ weeks, x t-(γ-1)*24*7 is the data of the previous γ-1 week; X m is the monthly cycle characteristic x t-δ*24*30 is the data of the previous δ months, x t-(δ-1)*24*30 It is the data of the previous δ-1 month; Step 2: TCN network construction The TCN network consists of multiple residual block networks. Each residual block has two dilated causal convolutions. After each dilated causal convolution, the weight is normalized, and the ReLU nonlinear activation function f(x) = max(0, x) and Dropout regularization are used. At the same time, a residual connection is performed every two layers of dilated causal convolutions. Step 3: MLP network construction An MLP network consists of two layers. The first layer consists of two linear layers, and the second layer is a linear layer network: In the formula, It is element-by-element multiplication, SiLU is a nonlinear activation function, Linear is a linear layer function, and the formulas of SiLU and Linear are as follows: In the formula, W is the weight parameter, b is the bias vector parameter, and the sigmoid formula is as follows: Step 4: Sigmoid gated network construction The Sigmoid network controls the weight of the previous network layer's contribution to the next network layer. The Sigmoid gating branch controls the weight of the TCN network. The Sigmoid gating branch consists of an MLP network and a Sigmoid activation function: Where G(x) is the gating network, It is element-by-element multiplication, Sigmoid is the activation function, MLP and TCN correspond to MLP network and TCN network respectively; Step 5: Construction of multi-expert load forecasting model The multi-expert load forecasting model is composed of a TCN network, an MLP network, and a Sigmoid gating network. Four different TCN networks receive the hourly cycle feature X h , daily cycle characteristics X d , Weekly characteristics X w 、Monthly cycle characteristics X m After the input, the outputs of the four TCN networks are obtained. After the outputs of the TCN networks pass through their respective Sigmoid gating networks, the gated network outputs corresponding to the hourly periodic features, daily periodic features, weekly periodic features, and monthly periodic features are obtained. The four groups of gated network outputs are spliced ​​together to obtain the gated network output; after the gated network output passes through a linear layer and softmax calculation, the topk expert index is selected according to the softmax result, and the selected MLP network is selected according to the index to participate in the calculation, and the gated network output is passed as input to the selected MLP network, and the weight calculated by the above softmax is accumulated and then accumulated to obtain the output of the multi-expert network; During the training data preparation process, the input data was divided into four groups: hourly cycle, daily cycle, weekly cycle, and monthly cycle. Each group of input data passed through a different TCN network and Sigmoid gating network, and the outputs of the four groups of gating networks were as follows: In the formula, x h is the hourly period input, x d is the daily cycle data, x w is the weekly periodic data, x m is the monthly cycle data, G h , G d , G w , G m is the gating network output corresponding to the periodic data; After the gating network output results of the four sets of periodic features are spliced ​​together, they pass through different MLP expert networks. Each MLP expert network models different input features, and only the topk experts are selected to participate in the network calculation to obtain the final prediction result: G=concat(G h ,G d ,G w ,G m ) weight,index=topk(softmax(WG+b)) In the formula, concat is the matrix concatenation calculation, W is the weight parameter of the linear layer, b is the bias vector of the linear layer, topk is the largest k numbers in the softmax calculation result, the largest k numerical results are represented by the weight vector, and the corresponding index value is represented by the index vector. The value of index is the selected expert index. The softmax calculation formula is as follows: In the formula, x is a set of vectors, x i is the i-th value in the vector; After calculating the selected expert index and the corresponding weight, use the selected expert MLP network to perform calculations, and sum the results of each expert network to get the final result: In the formula, y is the output of the multi-expert model, index is the index result of topk, weight is the weight value corresponding to topk, G is the output of the gating network of the previous layer, and MLP i is the i-th expert network; Step 6: Multi-expert model training and evaluation The Pytorch framework is used for model training and prediction, and the optimal model is trained through the gradient descent method.

2. A method for medium- and short-term power load forecasting based on a multi-expert model as claimed in claim 1, characterized in that: The TCN network is replaced by an LSTM or CNN network.

3. The method for medium- and short-term power load forecasting based on a multi-expert model as claimed in claim 1, characterized in that: Step 6 The error function for multi-expert model training and evaluation uses root mean square error RMSE and mean absolute error MAE. The error function formula is as follows: Where N is the number of samples, y i is the actual load value, is the predicted load value of the model.

Images