Short-term power load multi-step prediction method based on MCNN-MMoL

Abstract

The invention discloses a short-term power load multi-step prediction method based on MCNN-MMoL, and the method comprises the following steps: S101, collecting original load data, and carrying out the preprocessing of the original load data; s102, dividing the original load data into a plurality of sub-loads with independent characteristic information; and S103, constructing an MCNN-MMoL network, capturing and extracting the characteristics of the sub-load from different scales based on an MCNN part in the MCNN-MMoL network, carrying out multi-step load prediction based on an MMoL part in the MCNN-MMoL network, and finally obtaining a short-term power load multi-step prediction result. The method is used for solving the technical problem that existing power load prediction is single-step prediction and has large use limitation.

Description

Technical Field

[0001] The present application relates to the technical field of power load forecasting, and in particular to a short-term power load multi-step forecasting method based on MCNN-MMoL. Background Art

[0002] The degree of clean electricity in my country continues to improve, and the proportion of electricity consumption in the tertiary industry and residents' lives continues to increase, which further increases the uncertainty on both sides of the system source and load. The "Electricity Market Supervision Measures" implemented on June 1, 2023 clearly added virtual power plants as electricity trading entities, which will provide opportunities for flexible resources such as controllable loads to enter the market, and fully stimulate and release the flexible adjustment capabilities on the user side. Virtual power plants achieve joint management and optimized scheduling of decentralized power sources, loads and energy storage facilities through the energy Internet. Its operating principles mainly include: power trading, energy storage scheduling, and load scheduling. At the load scheduling level, virtual power plants with enhanced load regulation capabilities are an effective means to promote the development of demand response business, starting from "realizing accurate control of adjustable loads on the user side and improving users' enthusiasm for participating in demand response." Accurate load results help to tap the potential for accurate control of adjustable resources on the demand side, improve the accuracy of demand-side adjustable load potential assessment and market participation, and provide technical support for lean demand response construction and increased renewable energy penetration in the context of new power systems.

[0003] However, the existing power load forecasts are all single-step forecasts, which cannot be used in the formulation of power system day-ahead and day-intraday power generation plans of different time lengths, as well as in the actual application of power market bidding. Therefore, they have great limitations in use and cannot meet more needs.

[0004] Therefore, in order to solve the problems existing in the above-mentioned prior art, a short-term power load multi-step forecasting method based on MCNN-MMoL is proposed. Summary of the invention

[0005] The embodiment of the present application provides a short-term power load multi-step forecasting method based on MCNN-MMoL to solve the technical problem that the existing power load forecasts are all single-step forecasts, which cannot be used in the formulation of power system day-ahead and day-intraday power generation plans of different time lengths, as well as in the actual application of power market bidding, and have great limitations in use.

[0006] In view of this, the present application provides a short-term power load multi-step forecasting method based on MCNN-MMoL, comprising the following steps:

[0007] S101, collecting original load data, and preprocessing the original load data;

[0008] S102, dividing the original load data into a plurality of sub-loads having independent characteristic information;

[0009] S103, constructing an MCNN-MMoL network, capturing and extracting the features of the sub-load from different scales based on the MCNN part in the MCNN-MMoL network, performing multi-step load forecasting based on the MMoL part in the MCNN-MMoL network, and finally obtaining a short-term power load multi-step forecasting result.

[0010] Optionally, in step S101, the original load data is preprocessed, and the preprocessing specifically includes: identifying abnormal values ​​in the original load data, correcting the abnormal values ​​using a linear regression method, and performing normalization processing based on the correction results.

[0011] Optionally, in step S102, the MCNN part in the MCNN-MMoL network includes different convolution branches, and the different convolution branches are used to process the sub-loads at different levels.

[0012] Optionally, when the MMoL part in the MCNN-MMoL network performs multi-step load forecasting, it specifically includes: the multi-step load forecasting includes using multiple groups of long short-term memory networks to analyze the change law of the sub-load characteristics, obtain a prediction sequence, and determine the number of tasks and the number of gating units to be predicted according to the length of the prediction sequence, and customize a special gating network for each task, adjust the shared feature expression of each long short-term memory network for a specific task, and then perform nonlinear transformation on the output of the gating unit, and finally generate a prediction result through a linear layer;

[0013] The shared feature expression of the specific task is:

[0014]

[0015] Among them, Share i (*) is the shared feature representation of the i-th task, Lstm is the m-th shared subnetwork, and m = {1, 2, ..., Ls}, Gate i (*) m is the weight of the mth shared subnetwork in task i;

[0016] The expression of the gating network is:

[0017] Gate i (X t ″′)=softmax(X t ″′)

[0018] Among them, Gate i(*) is the gated network, Gate i (*) m is the weight of the mth shared subnetwork in task i, and softmax is the activation function;

[0019] The task output is:

[0020] y i =Tower i (Share i (X t ″′))

[0021] Among them, Share i (*)Tower i (*),y i , are the shared feature representation, tower network, and task output of the i-th task, respectively.

[0022] Optionally, the test set data is input into the prediction model with fixed parameters to obtain multi-step load prediction results, and RMSE, MAPE and R2 evaluation indicators are used to evaluate and analyze the performance of the prediction model.

[0023] Optionally, the RMSE, MAPE and R2 evaluation indicator formulas are:

[0024]

[0025] Among them, Len is the length of the sample set; y i and are the load mean, true value and predicted value corresponding to the i-th time point respectively.

[0026] Optionally, the relationship between the original load data and the sub-load data is expressed by an additive model function, and the additive model function can be expressed as:

[0027]

[0028] in, is the original load at time t, is the jth sub-load at that moment, the number of sub-loads is Le, and Le≥2.

[0029] Optionally, the expression for constructing the multi-level input sample in step S101 is:

[0030]

[0031] Among them, sample t is (X t ′,Y t ), the sample set is The sliding window size is p, is the branch input corresponding to the total load, It is the branch input corresponding to the jth sub-load.

[0032] Optionally, the step S102 includes, on the basis of constructing the multi-level input sample, performing convolution operation on the corresponding input load data of each branch to obtain a load output feature map, and using a concatenation layer to concatenate the load output feature map to form a feature representation as the input of the next layer of network. The relevant expression is:

[0033]

[0034] Among them, θ node ,θ i are the branch parameters corresponding to the original load and sub-load, respectively, t node″ , X″ t,i are the outputs of the original load and sub-load, respectively, Conv1D * (*) is a one-dimensional convolution operation, and Concatnate is a concatenation function.

[0035] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0036] The short-term power load multi-step forecasting method based on MCNN-MMoL provided in this application organically combines the multi-branch feature extraction architecture and the multi-task learning mechanism, provides more accurate and stable forecasting results, and can be used in the formulation of power system day-ahead and intraday power generation plans of different time lengths, as well as in the actual application of power market bidding, to meet more usage needs;

[0037] The multi-step prediction method based on multi-gated shared learning can dynamically map feature information to different tasks and improve model performance by jointly optimizing multiple tasks;

[0038] The multi-branch parallel CNN feature extraction network divides the feature space and decomposes the high-dimensional feature space optimization into multiple low-dimensional space sub-optimizations, thereby improving the comprehensiveness and meticulousness of feature extraction.

[0039] The multi-step prediction model built based on MCNN and MMoL can consider multi-scale feature information and differentially assign feature weights for prediction tasks, effectively improving the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly express the technical solutions of the embodiments of the present application, the drawings required for describing the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0041] Figure 1 A step diagram of a short-term power load multi-step forecasting method based on MCNN-MMoL provided in an embodiment of the present application;

[0042] Figure 2 A schematic diagram of the structure of the MCNN of the short-term power load multi-step forecasting method based on MCNN-MMoL provided in an embodiment of the present application;

[0043] Figure 3 A schematic diagram of the structure of MMoL of the short-term power load multi-step forecasting method based on MCNN-MMoL provided in an embodiment of the present application;

[0044] Figure 4 A comparison diagram of load curves at different prediction time steps under working day and non-working day conditions of the short-term power load multi-step forecasting method based on MCNN-MMoL provided in an embodiment of the present application;

[0045] Figure 5 This is a comparison chart of prediction curves of different models under working day and non-working day conditions of the short-term power load multi-step forecasting method based on MCNN-MMoL provided in an embodiment of the present application. DETAILED DESCRIPTION

[0046] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0047] For easier understanding, see Figures 1 to 3 The present invention provides a short-term power load multi-step forecasting method based on MCNN-MMoL, comprising the following steps:

[0048] S101, collecting original load data, and preprocessing the original load data;

[0049] S102, dividing the original load data into a plurality of sub-loads having independent characteristic information;

[0050] S103, constructing an MCNN-MMoL network, capturing and extracting the features of the sub-load from different scales based on the MCNN part in the MCNN-MMoL network, performing multi-step load forecasting based on the MMoL part in the MCNN-MMoL network, and finally obtaining a short-term power load multi-step forecasting result.

[0051] Specifically, step S101 includes identifying outliers in the original load data, and correcting the outliers using a linear regression method, and then selecting the features of the input original load data and normalizing them, and dividing the normalized features into a training set, a validation set, and a test set; step S102 includes hierarchical division and periodic encoding of the original load data, constructing multi-scale internal features and external features, and then using different convolution branches to process loads at different levels, and using a concatenation function to fuse multiple load data, and constructing a high-dimensional spatial feature matrix; step S103 includes performing hierarchical division and periodic encoding of the original load data, and constructing a multi-scale internal feature and external feature matrix. In 03, multiple groups of LSTMs are used to analyze the load variation pattern in the feature matrix, and the number of subtasks and gated units is determined according to the length of the prediction sequence. A special gated network is customized for each task, and the shared feature expression of each LSTM for a specific task is adjusted. The output of the gated unit is then nonlinearly transformed using an activation function, and finally the prediction result is generated through a linear layer to obtain the MCNN-MMoL model. Based on the training set, the network weights are adaptively adjusted through the Adam optimizer until the MCNN-MMoL model reaches a convergence state, completing the adjustment of the MCNN-MMoL model parameters.

[0052] Specifically, the Multi-gate Mixture of Experts (MMoE) is a learning machine composed of multiple shared components and task-specific gating units. The difference from the traditional multi-task learning machine is that MMoE can share knowledge between different tasks while maintaining the independence between tasks. Even if the correlation between some tasks is weak, MMoE can reduce the risk of negative transfer.

[0053] LSTM is good at processing periodic sequence data. It can not only capture complex time dynamics, but also has strong generalization ability. Therefore, LSTM is used as a shared module of MMoE to build a multi-task model MMoL that combines LSTM and multi-gating units.

[0054] When the prediction sequence length is q, MMoL consists of Ls shared subnetworks, q gated networks and tower networks. The number of gated networks and tower networks depends on the number of tasks. In MMoL, each shared subnetwork is an independent LSTM network responsible for capturing different dependencies from the feature fusion matrix output by MCNN.

[0055] The gated network also takes the feature fusion matrix as input, generates a probability vector after being processed by the activation function (softmax), and assigns weights to all shared subnetworks. Each task has a separate gate. These gate units selectively utilize different shared knowledge to generate the most relevant feature representation for each task. The tower network is the top network of each task, which is used to receive the weighted output of the corresponding gate unit and the bottom network, and generate the target task result.

[0056] The shared feature expression for a specific task is:

[0057]

[0058] Among them, Share i (*) is the shared feature representation of the i-th task, Lstm is the m-th shared subnetwork, and m = {1, 2, ..., Ls}, Gate i (*) m is the weight of the mth shared subnetwork in task i;

[0059] The expression of the gating network is:

[0060] Gate i (X t ″′)=softmax(X t ″′)

[0061] Among them, Gate i (*) is the gated network, Gate i (*) m is the weight of the mth shared subnetwork in task i, and softmax is the activation function;

[0062] The task output is:

[0063] y i =Tower i (Share i (X t ″′))

[0064] Among them, Share i (*), Tower i (*),y i They are the shared feature representation, tower network, and task output of the i-th task, respectively.

[0065] Furthermore, the test set data is input into the MCNN-MMoL model with fixed parameters to obtain the multi-step load forecasting results, and the RMSE, MAPE and R2 evaluation indicators are used to evaluate and analyze the performance of the MCNN-MMoL model.

[0066] Furthermore, the RMSE, MAPE and R2 evaluation index formulas are:

[0067]

[0068] Among them, Len is the length of the sample set; y i and are the load mean, true value and predicted value corresponding to the i-th time point respectively.

[0069] Furthermore, the relationship between the original load data and the sub-load data is expressed by an additive model function, which can be expressed as:

[0070]

[0071] in, is the original load at time t, is the jth sub-load at that moment, the number of sub-loads is Le, and Le≥2.

[0072] Furthermore, the expression for constructing the multi-level input sample in step S101 is:

[0073]

[0074] Among them, sample t is (X t ′,Y t ), the sample set is The sliding window size is p , is the branch input corresponding to the total load, It is the branch input corresponding to the jth sub-load.

[0075] Furthermore, step S102 includes, on the basis of constructing a multi-level input sample, performing convolution operation on the corresponding input load data of each branch to obtain a load output feature map, and using a concatenation layer to concatenate the load output feature map to form a more comprehensive and rich feature representation as the input of the next layer of network. The relevant expression is:

[0076]

[0077] Among them, θ node ,θ i are the branch parameters corresponding to the original load and sub-load, respectively, t node″ , X″ t,i are the outputs of the original load and sub-load, respectively, Conv1D * (*) is a one-dimensional convolution operation, and Concatnate is a concatenation function.

[0078] Furthermore, the reasonable selection of hyperparameters is an important way to ensure the accuracy of load forecasting. The input window size of MCNN-MMoL is 48, the number of features is 12, and the output dimension is the predicted sequence length. The model consists of two networks, in which the number of branches in the MCNN layer is 8 and the network type is CNN; the number of shared subnets in the MMoL layer is 9, the network type is LSTM, the number of gated units is equal to the predicted sequence length, the network type is linear, and the number of tower networks is also equal to the predicted length, and the network type is linear; the batch size is 1024, the number of iterations is 40, the optimizer is Adam, and the loss function is MAE. For other parameters, see Table 1 for details.

[0079]

[0080]

[0081] Table 1 MCNN-MMoL model hyperparameters

[0082] The performance of the proposed model is tested through three sets of experiments. First, different input scenarios are designed to analyze the importance of sample construction methods to the model; second, by comparing with other models, the superiority and effectiveness of the proposed model are verified by comparing the prediction results at different time steps; third, the necessity of each component module of the proposed model is verified.

[0083] First, the prediction results under different input conditions are compared.

[0084] In order to verify the adaptability of the sample construction method and the proposed model, the following four input scenarios are set for comparison:

[0085] Scenario 1: Without considering external auxiliary features, a CNN feature extraction network is constructed for the total load and each area load respectively;

[0086] Scenario 2: Considering external auxiliary features, a CNN feature extraction network is constructed for the total load, the load of each area, and each external feature;

[0087] Scenario 3: Considering external auxiliary features, we construct a number of branches equal to the number of features. The input of each CNN feature extraction network includes all feature variables.

[0088] Scenario 4: Considering external auxiliary features, a CNN feature extraction network is constructed for the total load and the load in each area respectively, and the external features are incorporated into the input of the branch corresponding to the total load.

[0089] Table 2 shows the prediction results in different scenarios when the prediction step number q is 24, 48, and 96 respectively. As can be seen from Table 2, the model effect is the best under scenario 4, and the worst under scenario 1 without considering external auxiliary features. This shows that there is a correlation between time features such as hours and days and loads, which helps to capture the regularity of load changes more comprehensively. The prediction error of scenario 3 is greater than that of scenario 2. This is because in scenario 2, these features are processed separately, resulting in the model being unable to fully utilize the time correlation between time features and historical load features, resulting in a larger prediction error. Although scenario 3 considers both time and load features, since the total load dominates the input data, the model may focus more on learning the fluctuations of the total load during the optimization process and ignore the key change information in the sub-load, thereby affecting the model's full utilization of load data. Scenario 4 considers the combination of time features and load features, and avoids the problem of total load dominance in scenario 3, thereby making full use of the hierarchical structure and time features of load data, and ultimately achieving the best prediction effect.

[0090]

[0091] Table 2 Comparison of prediction results under different input scenarios

[0092] Second, compare the prediction results with those of other models.

[0093] In order to more fully demonstrate the advantages of MCNN-MMoL in multi-step prediction, two sets of comparative experiments were conducted. First, MMoL was compared with LSTM based on direct prediction (D-LSTM), LSTM based on recursive prediction (R-LSTM), and LSTM based on Seq2Seq (S-LSTM) to verify the effectiveness of multi-step prediction based on multi-task architecture. Then, MCNN-MMoL was compared with CNN-LSTM, MCNN-LSTM, and MMoE models to verify the superiority of the proposed multi-step model. The batch size, number of iterations, optimizer, and loss function of all the compared models are the same as those of the proposed model. The number of D-LSTM models is equal to the number of prediction steps, and the number of branches of MCNN in MCNN-LSTM is also consistent with the proposed model.

[0094] Table 3 compares the evaluation results of different prediction strategies at different prediction steps. Figure 4The load curves at different prediction time steps under working days and non-working days are compared. The experimental results show that the prediction errors of all models increase with the increase of prediction time steps, and the error increase of R-LSTM is particularly obvious. This is because the recursive prediction method depends on the previous prediction value, and the error will continue to accumulate in each recursive step, that is, the more prediction steps, the greater the accumulated error. Another model with a significant decrease in prediction accuracy is S-LSTM. Although the decoder of S-LSTM uses teacher forcing during the training process, that is, the input of the decoder is the actual output value, the real output value is not available in the inference stage, so the model's own prediction value needs to be used as the input of the next step. The prediction error accumulates with iterations, thus affecting the accuracy of the prediction results. Although the error of MMoL also increases with the increase of time steps, its prediction curve fitting effect is better than the comparison model, indicating that the multi-task architecture can improve the prediction performance by utilizing the dependency information between tasks and jointly optimizing multiple related tasks.

[0095] As the time step increases, the prediction accuracy of D-LSTM and MMoL gradually approaches. In addition, the R2 index of D-LSTM and MMoL from 24 to 96 steps decreases by 5.23% and 7.48% respectively. D-LSTM performs more stably in long time step prediction. The reason may be that D-LSTM is essentially composed of multiple independent LSTM single-step prediction models. Each model is directly optimized for a specific time step and is not affected by other tasks. However, D-LSTM does not consider the position information between future time steps, resulting in discontinuous prediction curves and relatively large fluctuations. In general, MMoL has better performance.

[0096]

[0097] Table 3 Comparison of results of different prediction strategies

[0098] To further demonstrate the performance of the proposed model in multi-step prediction and compare it with other advanced models, Table 4 is the error comparison of different multi-step prediction models under different prediction steps. Figure 5 This is a comparison of prediction curves of different models, date selection and Figure 4The experimental results show that MCNN-MMoL outperforms other models in all evaluations under the same prediction step length, and the prediction curve has a higher degree of fitting. Especially in areas with large load fluctuations such as peaks and troughs, the model can capture the load change trend more accurately. Compared with CNN-LSTM, MCNN-LSTM with added convolution branches has better results, with MAPE reduced by 0.11%, 10.12%, and 10.46%, respectively, and RMSE reduced by 10.92kW, 15.53kW, and 34.46kW, respectively. R2 increased by 1.09%, 1.85%, and 5.01%, respectively. It can be seen that the modeling ability of CNN-LSTM in the 96-step prediction is significantly reduced, which may be due to the insufficient feature extraction ability of the simple CNN, which cannot fully characterize the changes in longer time steps. MCNN-LSTM can fully extract the potential regular information in the input data through a multi-branch convolutional network, thereby improving the long-time step prediction accuracy of the model. The R2 index of CNN-LSTM, MCNN-LSTM, Seq2Seq, MMoE, and MCNN-MMoL from 24 to 96 steps decreased by 15.86%, 11.94%, 26.78%, 7.70%, and 8.19%, respectively. The decrease of MMoE and MCNN-MMoL is relatively small, indicating that the stability of the two is strong. Although the prediction error of all models increases with the increase of time step, compared with other models, MCNN-MMoL provides more accurate and stable results for multi-step load forecasting with its effective feature extraction mechanism and multi-task prediction architecture.

[0099]

[0100] Table 4 Comparison of prediction results of different models

[0101] Third, it is necessary to verify the various components of the proposed model.

[0102] In order to verify the impact of the improved modules in the proposed model on the prediction results, two comparative models are designed. One is the CNN-MMoL model with only a single CNN network removed from the multi-branch architecture, and the other is the MCNN-MMoE model with the LSTM network removed. Table 5 shows the corresponding experimental results. It can be seen from the results that the errors of CNN-MMoL and MCNN-MMoE are higher than those of the proposed model, which proves the modeling ability of LSTM for time correlation and the advantages of the multi-branch CNN architecture in multi-level load data feature extraction.

[0103]

[0104] Table 5 Prediction results of ablation experiments

[0105] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein, for example. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0106] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A short-term power load multi-step forecasting method based on MCNN-MMoL is characterized by: The following steps are involved: S101, collecting original load data, and preprocessing the original load data; S102, dividing the original load data into a plurality of sub-loads having independent characteristic information; S103, constructing an MCNN-MMoL network, capturing and extracting the features of the sub-load from different scales based on the MCNN part in the MCNN-MMoL network, performing multi-step load forecasting based on the MMoL part in the MCNN-MMoL network, and finally obtaining a short-term power load multi-step forecasting result.

2. The short-term power load multi-step forecasting method based on MCNN-MMoL according to claim 1 is characterized in that: In the step S101, the original load data is preprocessed, and the preprocessing specifically includes: identifying abnormal values ​​in the original load data, correcting the abnormal values ​​using a linear regression method, and performing normalization processing based on the correction results.

3. The short-term power load multi-step forecasting method based on MCNN-MMoL according to claim 2 is characterized in that: In step S102, the MCNN part in the MCNN-MMoL network includes different convolution branches, and the different convolution branches are used to process the sub-loads at different levels.

4. The short-term power load multi-step forecasting method based on MCNN-MMoL according to claim 3 is characterized in that: When the MMoL part in the MCNN-MMoL network performs multi-step load forecasting, it specifically includes: using multiple groups of long short-term memory networks to analyze the change rules of the sub-load characteristics, deriving a prediction sequence, and determining the number of tasks and the number of gating units to be predicted according to the length of the prediction sequence, and customizing a special gating network for each task, adjusting the shared feature expression of each long short-term memory network for a specific task, and then performing nonlinear transformation on the output of the gating unit, and finally generating a prediction result through a linear layer; The shared feature expression of the specific task is: Among them, Share i (*) is the shared feature representation of the i-th task, Lstm is the m-th shared subnet, and m = {1, 2, ..., Ls}rGate i (*) m is the weight of the mth shared subnetwork in task i; The expression of the gating network is: Gate i (X t ″′)=softmax(X t ″′) Among them, Gate i (*) is the gated network, Gate i (*) m is the weight of the mth shared subnetwork in task i, and softmax is the activation function; The task output is: y i =Tower i (Share i (X t ″′)) Among them, Share i (*)Tower i (*),y i , are the shared feature representation, tower network, and task output of the i-th task, respectively.

5. The short-term power load multi-step forecasting method based on MCNN-MMoL according to claim 2 is characterized in that: The method includes inputting the test set data into the prediction model with fixed parameters to obtain multi-step load prediction results, and using RMSE, MAPE and R2 evaluation indicators to evaluate and analyze the performance of the prediction model.

6. The short-term power load multi-step forecasting method based on MCNN-MMoL according to claim 5 is characterized in that: The RMSE, MAPE and R2 evaluation index formulas are: Among them, Len is the length of the sample set; y i and are the load mean, true value and predicted value corresponding to the i-th time point respectively.

7. The short-term power load multi-step forecasting method based on MCNN-MMoL according to claim 1 is characterized in that: The relationship between the original load data and the sub-load data is expressed by an additive model function, which can be expressed as: in, is the original load at time t, is the jth sub-load at that moment, the number of sub-loads is Le, and Le≥2.

8. The short-term power load multi-step forecasting method based on MCNN-MMoL according to claim 1 is characterized in that: The expression for constructing the multi-level input sample in step S101 is: Among them, sample t is (X′ t ,Y t ), the sample set is The sliding window size is p, is the branch input corresponding to the total load, It is the branch input corresponding to the jth sub-load.

9. The short-term power load multi-step forecasting method based on MCNN-MMoL according to claim 8 is characterized in that: The step S102 includes, on the basis of constructing the multi-level input samples, performing convolution operation on the corresponding input load data of each branch to obtain a load output feature map, and using a concatenation layer to concatenate the load output feature map to form a feature representation as the input of the next layer of network. The relevant expression is: Among them, θ node ,θ i are the branch parameters corresponding to the original load and sub-load, respectively, t node″ , X″ t,i are the outputs of the original load and sub-load, respectively, Conv1D * (*) is a one-dimensional convolution operation, and Concatnate is a concatenation function.

Images