Short-term power load multi-step prediction method based on mcnn-mmol

Abstract

The application discloses a short-term power load multi-step prediction method based on MCNN-MMoL, which comprises the following steps: S101, collecting original load data and pre-processing the original load data; S102, dividing the original load data into a plurality of sub-loads with independent characteristic information; S103, constructing an MCNN-MMoL network, capturing and extracting the characteristics of the sub-loads from different scales based on the MCNN part in the MCNN-MMoL network, performing multi-step load prediction based on the MMoL part in the MCNN-MMoL network, and finally obtaining a short-term power load multi-step prediction result; and the technical problem of the existing power load prediction being single-step prediction and having a large use limitation is solved.

Description

Technical Field

[0001] This 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 Technology

[0002] The increasing cleanliness of my country's electricity supply and the growing proportion of electricity consumption in the tertiary sector and residential use have further increased uncertainty on both the power source and load sides of the power system. The "Electricity Market Supervision Measures," implemented on June 1, 2023, explicitly recognizes virtual power plants as new entities in electricity trading. This provides opportunities for flexible resources such as controllable loads to enter the market, fully stimulating and releasing the flexible adjustment capabilities of the user side. Virtual power plants achieve joint management and optimized scheduling of distributed power sources, loads, and energy storage facilities through the energy internet. Their operating principles mainly include: electricity trading, energy storage scheduling, and load scheduling. At the load scheduling level, virtual power plants, with the aim of "achieving precise control of adjustable loads on the user side and improving user participation in demand response," are an effective means to promote the development of demand response business. Accurate load results help to tap the potential for precise 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 under the new power system framework.

[0003] Existing power load forecasts are all single-step forecasts, which cannot be used for the formulation of day-ahead and intraday power generation plans of different time periods, or for the actual application of power market bidding. Therefore, they have significant limitations and cannot meet more needs.

[0004] Therefore, in order to address the problems existing in the above-mentioned technologies, a short-term power load multi-step prediction method based on MCNN-MMoL is proposed to solve the problem. Summary of the Invention

[0005] This application provides a multi-step forecasting method for short-term power load based on MCNN-MMoL, which addresses the technical problem that existing power load forecasting methods are all single-step forecasts, which cannot be used in the formulation of day-ahead and intraday power generation plans of power systems with different time lengths, or in the actual application of power market bidding, and thus have significant limitations in use.

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

[0007] S101. Collect raw load data and preprocess the raw load data;

[0008] S102. Divide the original load data into multiple sub-loads with independent characteristic information;

[0009] S103. Construct an MCNN-MMoL network. Based on the MCNN part of the MCNN-MMoL network, capture and extract the features of the sub-loads from different scales. Perform multi-step load prediction based on the MMoL part of the MCNN-MMoL network to finally obtain the multi-step prediction results of short-term power load.

[0010] In step S101, the original load data is preprocessed. The preprocessing specifically includes: identifying outliers in the original load data, correcting the outliers using linear regression, and normalizing the data based on the correction results.

[0011] In step S102, the MCNN part of the MCNN-MMoL network includes different convolutional branches, which are used to process the sub-loads at different levels.

[0012] When the MMoL part of the MCNN-MMoL network performs multi-step load prediction, it specifically includes: the multi-step load prediction includes using multiple sets of long short-term memory networks to analyze the changing patterns of the sub-load features, obtaining a prediction sequence, determining the number of tasks to be predicted and the number of gating units based on the length of the prediction sequence, customizing a dedicated gating network for each task, adjusting the shared feature representation of each long short-term memory network for a specific task, performing a nonlinear transformation on the output of the gating unit, and finally generating the prediction result through a linear layer;

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

[0014]

[0015] in, Let i be the shared feature representation for the i-th task. m For the m-th shared subnet, and , Let m be the weight of the m-th shared subnet in task i;

[0016] The expression for the gated network is:

[0017]

[0018] in, For gating networks, Let m be the weight of the m-th shared subnet in task i. For activation functions;

[0019] The task output is:

[0020]

[0021] in, These represent the shared feature representation, the tower network, and the task output for the i-th task, respectively.

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

[0023] Optionally, the formulas for the RMSE, MAPE, and R2 evaluation indicators are as follows:

[0024]

[0025]

[0026]

[0027] Where Len is the length of the sample set; , and These represent the average load, actual load, and predicted load at the i-th time point, respectively.

[0028] Optionally, the relationship between the original load data and the sub-load data is expressed using an additive model function, which can be represented as:

[0029]

[0030] in, The original load at time t, This is the j-th sub-load at this moment, and the number of sub-loads is... .

[0031] Optionally, the expression for constructing multi-level input samples in step S101 is:

[0032]

[0033] Where, sample t is The sample set is The sliding window size is , For the branch input corresponding to the total load, This is the input for the branch corresponding to the j-th subload.

[0034] Optionally, step S102 includes, based on the construction of multi-level input samples, performing convolution operations on the corresponding input load data for each branch to obtain a load output feature map, and using a concatenation layer to concatenate the load output feature maps to form a feature representation as the input to the next layer of the network. The relevant expression is:

[0035]

[0036]

[0037]

[0038] in, These are the branch parameters corresponding to the original load and sub-load, respectively. , These are the outputs of the original load and the sub-load, respectively. (*) represents a one-dimensional convolution operation. This is a concatenation function.

[0039] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:

[0040] The short-term power load multi-step prediction method based on MCNN-MMoL provided in this application organically combines a multi-branch feature extraction architecture and a multi-task learning mechanism, providing more accurate and stable prediction results. It can be used in the formulation of day-ahead and intraday power generation plans of power systems with different time lengths, as well as in the practical application of power market bidding, to meet more usage needs.

[0041] Multi-step prediction methods based on multi-gated shared learning can dynamically map feature information to different tasks and improve model performance by jointly optimizing multiple tasks.

[0042] The CNN feature extraction network based on multi-branch parallelism improves the comprehensiveness and detail of feature extraction by dividing the feature space and optimizing the high-dimensional feature space into multiple low-dimensional sub-optimizations.

[0043] The multi-step prediction model based on MCNN and MMoL can consider multi-scale feature information and assign feature weights differently for prediction tasks, effectively improving prediction accuracy. Attached Figure Description

[0044] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart illustrating the steps of the short-term power load multi-step prediction method based on MCNN-MMoL provided in the embodiments of this application.

[0046] Figure 2 This is a schematic diagram of the structure of the MCNN based on the MCNN-MMoL short-term power load multi-step prediction method provided in the embodiments of this application;

[0047] Figure 3 This is a schematic diagram of the structure of the MMoL in the short-term power load multi-step prediction method based on MCNN-MMoL provided in the embodiments of this application;

[0048] Figure 4 This is a comparison chart of load curves for different prediction time steps under weekday and non-weekday conditions for the short-term power load multi-step prediction method based on MCNN-MMoL provided in the embodiments of this application.

[0049] Figure 5 This is a comparison chart of prediction curves of different models under weekday and non-weekday conditions for the short-term power load multi-step prediction method based on MCNN-MMoL provided in the embodiments of this application. Detailed Implementation

[0050] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0051] For easier understanding, please refer to Figures 1 to 3 This invention provides a short-term multi-step power load forecasting method based on MCNN-MMoL, comprising the following steps:

[0052] S101. Collect raw load data and preprocess the raw load data;

[0053] S102. Divide the original load data into multiple sub-loads with independent characteristic information;

[0054] S103. Construct an MCNN-MMoL network. Based on the MCNN part of the MCNN-MMoL network, capture and extract the features of the sub-loads from different scales. Perform multi-step load prediction based on the MMoL part of the MCNN-MMoL network to finally obtain the multi-step prediction results of short-term power load.

[0055] Specifically, step S101 includes identifying outliers in the original load data, correcting them using linear regression, selecting features from the input original load data, normalizing them, and dividing the normalized features into training, validation, and test sets; step S102 includes hierarchical partitioning and periodic encoding of the original load data, constructing multi-scale internal and external features, using different convolutional branches to process loads at different levels, and using a concatnate function to fuse the multi-path load data and construct a high-dimensional feature matrix; step S1 In step 03, multiple LSTMs are used to analyze the load variation patterns in the feature matrix. The number of sub-tasks and gating units are determined based on the length of the prediction sequence. A dedicated gating network is customized for each task. The shared feature representation of each LSTM for a specific task is adjusted. The output of the gating unit is then nonlinearly transformed using an activation function. 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 using the Adam optimizer until the MCNN-MMoL model reaches convergence, thus completing the adjustment of the MCNN-MMoL model parameters.

[0056] Specifically, a multi-gate mixture of experts (MMoE) is a learning machine composed of multiple shared components and task-specific gating units. Unlike traditional multi-task learning machines, 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.

[0057] LSTM excels at processing periodic sequence data, not only capturing complex temporal dynamics but also possessing strong generalization ability. Therefore, LSTM is used as a shared module of MMoE to build a multi-task model MMoL that combines LSTM and multiple gating units.

[0058] When the predicted sequence length is q, MMoL consists of Ls shared sub-networks, q gated networks and tower networks. The number of gated networks and tower networks depends on the number of tasks. In MMoL, each shared sub-network is an independent LSTM network responsible for capturing different dependencies from the feature fusion matrix output by MCNN.

[0059] The gated network also takes the feature fusion matrix as input, and generates a probability vector after being processed by the activation function (softmax). It 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, as the top-level network for each task, is used to receive the weighted output of the corresponding gate unit and the bottom network, and generate the target task result.

[0060] The shared feature expression for a specific task is:

[0061]

[0062] in, Let i be the shared feature representation for the i-th task. m For the m-th shared subnet, and , Let m be the weight of the m-th shared subnet in task i;

[0063] The expression for the gated network is:

[0064]

[0065] in, For gating networks, Let m be the weight of the m-th shared subnet in task i. For activation functions;

[0066] The task output is:

[0067]

[0068] in, , These represent the shared feature representation, the tower network, and the task output for the i-th task, respectively.

[0069] Furthermore, the test set data was input into the MCNN-MMoL model with fixed parameters to obtain multi-step load prediction results, and the performance of the MCNN-MMoL model was evaluated and analyzed using RMSE, MAPE and R2 evaluation metrics.

[0070] Furthermore, the formulas for RMSE, MAPE, and R2 evaluation indicators are as follows:

[0071]

[0072]

[0073]

[0074] Where Len is the length of the sample set; , and These represent the average load, actual load, and predicted load at the i-th time point, respectively.

[0075] Furthermore, the relationship between the original load data and the sub-load data is expressed using an additive model function, which can be represented as:

[0076]

[0077] in, The original load at time t, Let j be the j-th subload at this moment, and the number of subloads is . .

[0078] Furthermore, the expression for constructing multi-level input samples in step S101 is as follows:

[0079]

[0080] Where, sample t is The sample set is The sliding window size is , For the branch input corresponding to the total load, This is the input for the branch corresponding to the j-th subload.

[0081] Furthermore, step S102 includes, based on the construction of multi-level input samples, performing convolution operations on the corresponding input load data for each branch to obtain a load output feature map, and using a concatenation layer to concatenate the load output feature maps to form a more comprehensive and richer feature representation as the input to the next layer of the network. The relevant expression is:

[0082]

[0083]

[0084]

[0085] in, These are the branch parameters corresponding to the original load and sub-load, respectively. , These are the outputs of the original load and the sub-load, respectively. (*) represents a one-dimensional convolution operation. This is a concatenation function.

[0086] Furthermore, the appropriate 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 length of the predicted sequence. The model consists of two networks: the MCNN layer has 8 branches and is a CNN network; the MMoL layer has 9 shared subnetworks and is an LSTM network. The number of gated units is equal to the length of the predicted sequence and is a linear network. The number of tower networks is also equal to the length of the predicted sequence and is a linear network. The batch size is 1024, the number of iterations is 40, the optimizer is Adam, and the loss function is MAE. Other parameters are detailed in Table 1.

[0087]

[0088] Table 1. Hyperparameters of the MCNN-MMoL model

[0089] The performance of the proposed model is tested through three sets of experiments. First, different input scenarios are designed to analyze the importance of the sample construction method to the model. Second, the model is compared with other models to verify its superiority and effectiveness by comparing the prediction results at different time steps. Third, the necessity of each component module of the proposed model is verified.

[0090] First, a comparison of prediction results under different input conditions.

[0091] To verify the compatibility between the sample construction method and the proposed model, the following four input scenarios were set for comparison:

[0092] Scenario 1: Without considering external auxiliary features, construct a separate CNN feature extraction network for the total load and the load of each jurisdiction;

[0093] Scenario 2: Considering external auxiliary features, construct one CNN feature extraction network for the total load, the load of each jurisdiction, and each external feature;

[0094] Scenario 3: Considering external auxiliary features, construct a number of paths equal to the number of features, with the input of each CNN feature extraction network including all feature variables;

[0095] Scenario 4: Considering external auxiliary features, construct a CNN feature extraction network for the total load and the load of each jurisdiction, and incorporate the external features into the input of the corresponding branch of the total load.

[0096] Table 2 shows the prediction performance under different scenarios when the prediction steps q are 24, 48, and 96, respectively. As can be seen from Table 2, the model performs best under scenario 4, while scenario 1, which does not consider external auxiliary features, performs the worst. This indicates a correlation between hourly and daily time features and load, helping to more comprehensively capture the regularity of load changes. The prediction error of scenario 3 is greater than that of scenario 2 because in scenario 2, these features are processed separately, causing the model to fail to fully utilize the temporal correlation between time features and historical load features, resulting in a larger prediction error. Although scenario 3 considers both time and load features, because total load dominates the input data, the model may focus more on learning the fluctuations of total load during optimization, ignoring key changes in sub-loads, thus affecting the model's full utilization of load data. Scenario 4 considers both time and load features while avoiding the problem of total load dominance in scenario 3, thus fully utilizing the hierarchical structure and time features of load data, ultimately achieving the best prediction performance.

[0097]

[0098] Table 2 Comparison of prediction results under different input scenarios

[0099] Second, compare the prediction results with those of other models.

[0100] To more comprehensively 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 a 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 comparative models were the same as those of the proposed model. The number of D-LSTM models was equal to the number of prediction steps, and the number of branches of MCNN in MCNN-LSTM was also consistent with the proposed model.

[0101] Table 3 compares the evaluation results of different prediction strategies under different prediction steps. Figure 4The experiment compares the load curves for different prediction time steps under weekday and non-weekday conditions. The results show that the prediction error of all models increases with the number of prediction time steps, with R-LSTM showing a particularly significant increase. This is because recursive prediction methods rely on previous predictions, and the error accumulates with each recursive step; the more prediction steps, the greater the accumulated error. S-LSTM also shows a significant decrease in prediction accuracy. Although the S-LSTM decoder uses teacher coercion during training (i.e., the decoder input is the actual output value), the actual output value is unavailable during inference, so the model's own predictions are used as the input for the next step. The prediction error accumulates with iterations, thus affecting the accuracy of the prediction results. Although MMoL's error also increases with the number of time steps, its prediction curve fitting effect is generally better than the comparison models, indicating that multi-task architectures can improve prediction performance by utilizing the dependency information between tasks and jointly optimizing multiple related tasks.

[0102] As the time step increases, the prediction accuracy of D-LSTM and MMoL gradually converges. Furthermore, the R² prediction scores for D-LSTM and MMoL decrease by 5.23% and 7.48% respectively from 24 to 96 steps. D-LSTM performs more stably in longer prediction steps, possibly because it is essentially composed of multiple independent LSTM single-step prediction models, each optimized directly for a specific time step and unaffected by other tasks. However, D-LSTM does not consider positional information between future time steps, resulting in discontinuous prediction curves and relatively large fluctuations. Overall, MMoL exhibits better performance.

[0103]

[0104] Table 3 Comparison of Results for Different Prediction Strategies

[0105] To further demonstrate the performance of the proposed model in multi-step prediction and to compare it with other advanced models, Table 4 shows the error comparison of different multi-step prediction models under different prediction steps. Figure 5 This is a comparison of prediction curves from different models, and the choice of date. Figure 4Consistent results. Experimental results show that MCNN-MMoL outperforms other models in all evaluations under the same prediction step size, and has a higher degree of prediction curve fitting. Especially in regions with large load fluctuations such as peaks and troughs, the model can more accurately capture the trend of load changes. Compared with CNN-LSTM, MCNN-LSTM with added convolutional branches performs better, 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 is improved by 1.09%, 1.85%, and 5.01%, respectively. It can be seen that the modeling ability of CNN-LSTM is significantly reduced in 96 prediction steps, which may be because the feature extraction ability of CNN alone is insufficient and cannot fully represent the changes over a longer time step. MCNN-LSTM, through its multi-branch convolutional network, can fully extract the hidden patterns in the input data, thereby improving the model's long-time prediction accuracy. The R² prediction scores for CNN-LSTM, MCNN-LSTM, Seq2Seq, MMoE, and MCNN-MMoL decreased by 15.86%, 11.94%, 26.78%, 7.70%, and 8.19% respectively from 24 to 96 steps. MMoE and MCNN-MMoL showed relatively smaller decreases, indicating stronger stability. Although the prediction error of all models increases with time step size, compared to other models, MCNN-MMoL, with its effective feature extraction mechanism and multi-task prediction architecture, provides more accurate and stable results for multi-step load prediction.

[0106]

[0107] Table 4 Comparison of prediction results from different models

[0108] Third, verify the necessity of each component module of the proposed model.

[0109] To verify the impact of the improved modules in the proposed model on the prediction results, two comparative models were designed: one is the CNN-MMoL model, which removes the multi-branch architecture and contains only a single-path CNN; the other is the MCNN-MMoE model, which removes the LSTM network. Table 5 shows the corresponding experimental results. As can be seen from the results, the errors of both CNN-MMoL and MCNN-MMoE are higher than those of the proposed model, which proves the ability of LSTM to model temporal correlation and the advantages of the multi-branch CNN architecture in feature extraction of multi-level load data.

[0110]

[0111] Table 5. Predicted results of ablation experiments

[0112] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0113] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A short-term power load multi-step forecasting method based on MCNN-MMoL, characterized in that, Includes the following steps: S101. Collect raw load data and preprocess the raw load data; S102. Divide the original load data into multiple sub-loads with independent characteristic information; S103. Construct an MCNN-MMoL network. Based on the MCNN part of the MCNN-MMoL network, capture and extract the features of the sub-loads from different scales. Perform multi-step load prediction based on the MMoL part of the MCNN-MMoL network to finally obtain the multi-step prediction results of short-term power load. In step S101, the original load data is preprocessed. The preprocessing specifically includes: identifying outliers in the original load data, correcting the outliers using linear regression, and normalizing the data based on the correction results. In step S102, the MCNN part of the MCNN-MMoL network includes different convolutional branches, which are used to process the sub-loads at different levels. When the MMoL part of the MCNN-MMoL network performs multi-step load prediction, it specifically includes: using multiple sets of long short-term memory networks to analyze the changing patterns of the sub-load features, obtaining a prediction sequence, determining the number of tasks to be predicted and the number of gating units based on the length of the prediction sequence, customizing a dedicated gating network for each task, adjusting the shared feature representation of each long short-term memory network for a specific task, performing a nonlinear transformation on the output of the gating unit, and finally generating the prediction result through a linear layer; The shared feature expression for the specific task is: in, Let i be the shared feature representation for the i-th task. m For the m-th shared subnet, and , Let m be the weight of the m-th shared subnet in task i; The expression for the gated network is: in, For gating networks, Let m be the weight of the m-th shared subnet in task i. For activation functions; The task output is: in, These represent the shared feature representation, the tower network, and the task output for the i-th task, respectively.

2. The short-term power load multi-step prediction method based on MCNN-MMoL according to claim 1, characterized in that, This includes inputting test set data into the MMoL part of the MCNN-MMoL network with fixed parameters to obtain multi-step load prediction results, and using RMSE, MAPE and R2 evaluation metrics to evaluate and analyze the performance of the MMoL part of the MCNN-MMoL network.

3. The short-term power load multi-step prediction method based on MCNN-MMoL according to claim 2, characterized in that, The formulas for the RMSE, MAPE, and R2 evaluation indicators are as follows: Where Len is the length of the sample set; , and These represent the average load, actual load, and predicted load at the i-th time point, respectively.

4. The short-term power load multi-step prediction method based on MCNN-MMoL according to claim 1, 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 represented as: in, The original load at time t, This is the j-th sub-load at this moment, and the number of sub-loads is... .

5. The short-term power load multi-step prediction method based on MCNN-MMoL according to claim 4, characterized in that, The expression for constructing multi-level input samples in step S101 is as follows: Where, the sample at time t is The sample set is The sliding window size is , For the branch input corresponding to the total load, This is the input for the branch corresponding to the j-th subload.

6. The short-term power load multi-step prediction method based on MCNN-MMoL according to claim 5, characterized in that, Step S102 includes, based on the construction of multi-level input samples, performing convolution operations on the corresponding input load data for each branch to obtain a load output feature map, and then using a concatenation layer to concatenate the load output feature maps to form a feature representation as the input to the next layer of the network. The relevant expression is: in, These are the branch parameters corresponding to the original load and sub-load, respectively. , The outputs of the original load and the sub-load are respectively. (*) represents a one-dimensional convolution operation. This is a concatenation function.

Images