Power supply insurance trend prediction method based on deep learning

Abstract

The invention discloses a deep learning-based power supply insurance trend prediction method, which comprises the following steps of: collecting historical power load time sequence data, and reconstructing and preprocessing the time sequence data; performing depth feature extraction on the processed time series data based on the trained CNN model; and based on the extracted depth features, carrying out multi-granularity decision making by using the trained depth forest, and obtaining power load prediction. According to the method, the Monte Carlo algorithm and the deep learning method are combined, various change scenes of the power load are simulated, quantitative evaluation of prediction result uncertainty is realized, a prediction mean value, a variance and a confidence interval can be provided, random fluctuation of the power load caused by influences of factors such as weather, holidays and festivals and emergencies can be better dealt with, and the prediction accuracy is improved. And a more comprehensive reference basis is provided for a power system scheduling decision.

Description

Technical Field

[0001] This invention relates to the field of smart grid power load forecasting technology, and in particular to a deep learning-based method for forecasting power supply trends. Background Technology

[0002] With the continuous growth of electricity demand and the increasing complexity of power grid operation, power load forecasting, as a key link in ensuring the stable operation of the power system, is facing increasingly greater challenges.

[0003] Most existing power load forecasting methods rely on physical models or traditional machine learning algorithms (such as convolutional neural networks, deep learning, Monte Carlo algorithms, and time series modeling methods using deep learning algorithms like Long Short-Term Memory (LSTM). While these methods can provide predictions to some extent, they still have many shortcomings and deficiencies when facing the uncertainty of power load fluctuations, including:

[0004] 1) While LSTM and other deep learning methods can effectively capture the temporal characteristics and trend changes of electricity load data, they have significant shortcomings in handling uncertainties in electricity load. Electricity load is affected by various external factors (such as weather, holidays, and sudden events), and these uncertainties are not fully incorporated into the prediction framework of existing models, resulting in a significant reduction in prediction accuracy when facing sudden fluctuations or unknown changes. 2) Short-term fluctuations in electricity load, especially under specific weather conditions or holidays, can exhibit extreme volatility. Although existing models can capture long-term load changes, their prediction accuracy for such short-term fluctuations is low, especially when data is scarce or changes are sudden, making the models prone to prediction bias. 3) Existing models typically rely on single time-series data for training, neglecting the impact of external features (such as weather data, social activities, and holidays) on electricity load. Although the prediction accuracy can be improved by introducing external features, the feature fusion capabilities of existing methods are limited, failing to fully explore the relationships between these multi-dimensional data.

[0005] For example, invention application No. 202510105966.X discloses a load forecasting method, device, and medium based on SPM-LSTM. This application, based on a subset of reference data and a reference LSTM model, can perform load forecasting for microgrids. However, its solution also has limitations. When dealing with complex and variable power load data, the generalization ability of the forecasting model needs improvement. Especially when facing load fluctuations caused by complex weather, sudden events, or large-scale social activities, the model may not be able to adapt quickly and provide accurate forecast results.

[0006] Therefore, existing models cannot effectively handle random fluctuations and sudden events in power load, resulting in large prediction errors, which in turn affect the accuracy and timeliness of power dispatch.

[0007] A deep learning-based power load forecasting method is needed. By introducing deep learning technology, especially considering the uncertainty of power fluctuations, the accuracy and reliability of power load forecasting can be significantly improved. Introducing a more flexible and intelligent deep learning model can effectively handle the randomness and uncertainty in power load, providing more accurate and reliable load forecasting results for the power system, thereby optimizing power system scheduling and operation management. Summary of the Invention

[0008] To address the aforementioned problems, the present invention aims to provide a deep learning-based method for predicting power supply trends, thereby solving the problem of insufficient generalization ability of existing load forecasting methods when dealing with complex and variable power load data.

[0009] This invention provides a deep learning-based method for predicting power supply trends.

[0010] The first aspect: a deep learning-based method for predicting power supply trends, including:

[0011] S1. Collect historical power load time-series data and perform time-series data reconstruction and preprocessing;

[0012] S2. Deep feature extraction is performed on the processed time-series data based on the trained CNN model;

[0013] S3. Based on the extracted deep features, multi-granularity decision-making is performed using the trained deep forest to obtain power load forecasts.

[0014] In one embodiment of the present invention, the time-series data reconstruction and preprocessing in S1 includes:

[0015] Reconstruct historical power load time series data into two-dimensional gridded data;

[0016] A similar working condition matching mechanism is constructed, and historical similar day data is filtered through the DTW algorithm.

[0017] In one embodiment of the present invention, in S2, the CNN model is based on the Monte Carlo dropout regularization method to quantify the uncertainty of the CNN model.

[0018] In one embodiment of the present invention, the uncertainty of a CNN model is quantified based on Monte Carlo methods, including:

[0019] During the training phase of CNN models, a dropout regularization mechanism is introduced into the fully connected layers to force the model to learn redundant representations.

[0020] During the prediction phase of the CNN model, multiple Monte Carlo samplings are performed, with different neurons being randomly activated during each forward propagation.

[0021] In one embodiment of the present invention, when performing multiple Monte Carlo samplings, the prediction results from the multiple samplings are used to calculate the statistical mean µ and variance. The uncertainty of CNN models is analyzed and quantified by the following formula:

[0022]

[0023]

[0024] Where T is the number of Monte Carlo samples, y t These are the predicted values ​​from a single sample, where µ is the mean. Let Variance be the variance.

[0025] In one embodiment of the present invention, the CNN model in S2 extracts deep features, including:

[0026] Based on a dual-channel convolutional structure, local temporal texture features are extracted and long-term trend features are captured.

[0027] Low-level features and high-level features are spliced ​​together by skip connections;

[0028] The pooling layer employs hybrid max-average pooling, which preserves key features while suppressing noise.

[0029] In one embodiment of the present invention, the deep forest multi-granularity decision-making process in S3 includes:

[0030] Based on the multi-granularity scanning layer, sliding window sampling is performed on the deep features output by the CNN to generate multi-scale feature fragments;

[0031] A combination of random forest and fully random forest is used to screen for high-discrimination features;

[0032] An attention-weighted mechanism is introduced to strengthen the decision-making weight of key features.

[0033] In one embodiment of the present invention, during deep forest training, an adaptive enhancement strategy is adopted to dynamically adjust the number of forest layers based on the prediction error.

[0034] A second aspect: an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, performs the steps of the method provided in the first aspect.

[0035] Third aspect: A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method provided in the first aspect.

[0036] The beneficial effects of this invention are:

[0037] 1. This invention innovatively combines the Monte Carlo algorithm with deep learning methods to quantitatively assess the uncertainty in prediction results by simulating various power load variation scenarios. Compared to traditional methods that only output a single, definitive value, this method provides the prediction mean, variance, and confidence interval (e.g., 95% confidence interval), clearly reflecting the fluctuation range of the prediction results. This characteristic enables it to better cope with random fluctuations in power load caused by factors such as weather, holidays, and emergencies, providing a more comprehensive reference for power system dispatching decisions, significantly improving the reliability of prediction results, and effectively addressing uncertainty.

[0038] 2. This invention overcomes the bottleneck of traditional models' insufficient ability to predict short-term fluctuations by introducing a more adaptable hybrid model structure. The dual-channel convolutional neural network employed can accurately capture short-term fluctuation patterns in power load. Combined with historical data filtered through a similar operating condition matching mechanism, it can provide more accurate short-term predictions even in scenarios with severe load fluctuations such as special weather or holidays. Simultaneously, the model can flexibly respond to the impact of sudden events on the load, reducing prediction bias caused by sudden changes, enhancing the ability to capture short-term fluctuations, and improving prediction accuracy in special scenarios.

[0039] 3. This invention achieves efficient fusion of multi-dimensional features through a collaborative architecture of Convolutional Neural Networks (CNN) and Deep Forest. CNN is responsible for extracting temporal features from power load data, while Deep Forest integrates external factors such as weather and holidays. It strengthens the decision weights of key features (such as the nonlinear relationship between temperature and load) through multi-granularity scanning and attention weighting mechanisms. This fusion method fully explores the potential correlations between multi-dimensional data, enabling the model to capture more comprehensive information on power load changes, significantly improving the overall accuracy and robustness of predictions, and comprehensively enhancing prediction accuracy. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating the principle of the intelligent grid power load accurate prediction method of the present invention.

[0041] Figure 2 This is a flowchart illustrating the intelligent grid power load accurate prediction method of the present invention.

[0042] Figure 3This is a schematic diagram of the structure of the smart grid power load accurate prediction model of the present invention;

[0043] Figure 4 This is a schematic diagram of the structure of the electronic device of the present invention. Detailed Implementation

[0044] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0045] Accurate power load forecasting is a crucial task in the operation and management of smart grids. Traditional forecasting methods often rely on historical data and statistical models, but their accuracy and adaptability are often limited when faced with complex and ever-changing power demand patterns.

[0046] To address the above problems, this invention provides a deep learning-based method for predicting power supply trends. Figure 1 The flowchart illustrates the principle of the intelligent grid power load accurate prediction method provided in this embodiment of the invention.

[0047] Example 1:

[0048] like Figure 2 As shown, this invention discloses a deep learning-based method for predicting power supply trends, including the following steps:

[0049] S1. Collect historical power load time-series data and perform time-series data reconstruction and preprocessing.

[0050] The process involves compiling sufficient time series of short-term power load data under similar operating conditions from historical data. Specific steps include:

[0051] Historical power load time series data (such as 15-minute granular load values, temperature, humidity, and other multi-dimensional indicators) are reconstructed into two-dimensional gridded data, including:

[0052] In the time dimension, a sliding window is used to divide the data into multiple time segments with a fixed-length time window. Each time segment contains historical load data within a certain time range.

[0053] In terms of feature dimensions, a multivariate channel is used, with different types of feature data mapped to different channels, similar to the RGB channels in an image. Channel 1 is for power load data; channel 2 is for temperature data; channel 3 is for humidity data; and channel 4 is for date type identifiers (e.g., 0 for weekdays and 1 for holidays). After the above processing, the data in each sliding window forms a two-dimensional matrix, where the rows of the matrix represent time steps and the columns represent different feature channels.

[0054] This processing method transforms the original time-series data into image-like data, facilitating feature extraction and learning using image processing techniques such as Convolutional Neural Networks (CNNs) in deep learning. This approach helps the model capture the temporal characteristics of load changes, providing a foundation for subsequent deep learning model training.

[0055] Then, a similar working condition matching mechanism is constructed, using the DTW dynamic time warping algorithm to filter historical similar day data, thereby enhancing the consistency of working conditions among training samples. The content includes:

[0056] Features that have a significant impact on power load, such as temperature, load change trends, and date type, are selected as inputs to the DTW algorithm. The DTW distance between the historical time period corresponding to the day to be predicted and the same time period of each historical day is calculated. The smaller the distance, the more similar the operating conditions of the two are. A similarity threshold is set, and historical day data with a DTW distance less than the threshold are selected as similar operating condition data and included in the training sample set to enhance the consistency of operating conditions of the training samples.

[0057] S2. Deep feature extraction is performed on the processed time-series data based on the trained CNN model. Specifically, this includes:

[0058] A dual-channel convolutional feature extraction mechanism is used. In channel 1, a 3×3 convolutional kernel is used to perform convolution operations on the input two-dimensional gridded data to extract local temporal texture features and capture short-term load fluctuation patterns. In channel 2, a 5×5 convolutional kernel is used to perform convolution processing to focus on capturing long-term trend features and revealing load change patterns on longer time scales such as daily and weekly.

[0059] Cross-layer feature fusion strategy. Features output from convolutional layer 1 (lower layer) and features output from convolutional layer 2 (higher layer) are concatenated through skip connections to integrate feature information from different layers and preserve multi-scale temporal patterns.

[0060] Hybrid pooling method. Hybrid max-average pooling is applied to the fused feature map. This downsamples the feature map to reduce data dimensionality and improve computational efficiency, while max pooling preserves key features and average pooling suppresses noise interference, ultimately yielding depth-extracted feature information that provides a solid foundation for subsequent processing.

[0061] This invention utilizes a deep convolutional neural network (CNN) model to extract hidden patterns in data through multiple layers of convolution and pooling operations. These patterns may not be easily characterized or discovered using traditional analytical methods. Convolutional kernels are used to extract texture features from input data (such as electricity load images), revealing complex temporal patterns and trends in electricity load data. Pooling layers reduce data dimensionality by downsampling feature maps, improving the model's computational efficiency while preserving the key information of the features. After processing through two convolutional and pooling layers, the deep convolutional network can extract more fundamental and richer feature information from the original data, laying a solid foundation for subsequent advanced inference and prediction.

[0062] S3. Based on the extracted deep features, multi-granularity decision-making is performed using the trained deep forest to obtain power load forecasts.

[0063] Deep forest is an ensemble learning method based on decision trees that can effectively process and classify input features. Through training, a deep forest model is built that can predict power load based on extracted features, ultimately achieving accurate predictions of power load for future forecast periods.

[0064] When using deep forests for multi-granularity decision-making, the following are included:

[0065] Based on multi-granularity scanning layers, sliding window sampling (e.g., window size 100 / 200 / 400) is applied to the depth features output by the CNN to generate multi-scale feature fragments. These multi-scale feature fragments contain load variation information at different time granularities, which helps to capture the complex variation patterns of power load over different time periods. By analyzing these feature fragments, a more comprehensive understanding of the dynamic characteristics of power load can be achieved.

[0066] Deep forests employ a cascaded forest structure, which combines random forests with fully random forests to select features with high discriminative power; and introduces an attention weighting mechanism to strengthen the decision weight of key features (such as the nonlinear relationship between temperature and load).

[0067] By cascading multiple forest layers, each layer can further refine and enhance predictive capabilities based on the features output by the previous layer. Random forests effectively reduce the risk of overfitting and improve the model's generalization performance by randomly selecting features and sample subsets to construct decision trees. Fully random forests further relax the randomness of feature selection, enhancing the model's robustness to noise. The combined use of these two methods can more comprehensively mine and utilize information from deep features, filtering out high-discriminative features that have a significant impact on prediction results.

[0068] The introduction of attention weighting is intended to more accurately capture features crucial for electricity load forecasting. In smart grids, certain features (such as the nonlinear relationship between temperature and load) can significantly impact forecast results. Through attention weighting, the weight of these key features in the decision-making process is increased, allowing the model to focus more intently on these important features, thereby improving the accuracy and stability of forecasts.

[0069] Furthermore, during deep forest training, an adaptive enhancement strategy is adopted to dynamically adjust the number of forest layers (up to 10 layers) based on the prediction error, thereby avoiding overfitting.

[0070] In the early stages of training, when the prediction error is large, the number of forest layers is increased, allowing the model to learn more complex features from the data. As training progresses and the prediction error gradually decreases, the number of forest layers is appropriately reduced to prevent the model from learning noisy data and to maintain the model's simplicity and generalization ability. This dynamic adjustment mechanism enables the model to adaptively optimize its structure according to the training progress at different stages, ensuring that while maintaining prediction accuracy, computational resources and time consumption are minimized.

[0071] Example 2:

[0072] This embodiment discloses a power supply trend prediction method based on deep learning. Based on embodiment 1, the uncertainty of the CNN model is quantified by applying the Monte Carlo dropout regularization method.

[0073] Monte Carlo dropout regularization is a method that simulates model uncertainty by randomly dropping neurons in a network. Introducing Monte Carlo dropout regularization during CNN model training can quantify the uncertainty of model parameters, thereby improving the stability and accuracy of predictions. This method randomly selects a subset of neurons to drop in each iteration, preventing the model from becoming overly reliant on certain specific neurons during training, thus enhancing the model's generalization ability.

[0074] By introducing dropout regularization, the model randomly discards a portion of neurons during training, which helps prevent overfitting and allows the model to generalize better. This, in turn, better reflects the potential uncertainties in electricity load data.

[0075] Specifically, the uncertainty of CNN models is quantified based on Monte Carlo methods, including:

[0076] During the training phase of a CNN model, a dropout regularization mechanism (e.g., a dropout rate of 0.5) is introduced into the fully connected layers to force the model to learn redundant representations. After applying dropout regularization to the fully connected layers, each neuron has a certain probability of being randomly dropped in each training iteration. This means that the network cannot rely on any single neuron for prediction. This mechanism encourages the network to learn more distributed feature representations, so that even if some neurons are dropped, other neurons can take over their work, maintaining the overall performance of the model.

[0077] During the prediction phase of the CNN model, multiple Monte Carlo samplings are performed, with different neurons being randomly activated during each forward propagation.

[0078] When performing multiple Monte Carlo samplings, the prediction results from these multiple samplings are used to statistically calculate the mean µ and variance. The uncertainty of CNN models is analyzed and quantified by the following formula:

[0079]

[0080]

[0081] Where T is the number of Monte Carlo samples, y t These are the predicted values ​​from a single sample, where µ is the mean. Let Variance be the variance.

[0082] Calculate the mean µ and variance of the prediction results Generate confidence intervals (e.g., 95% confidence level corresponds to...). )

[0083] The above calculations allow for the evaluation of the reliability of CNN model predictions, which can then be used to guide actual smart grid power load forecasting. A narrow confidence interval indicates stable model predictions and high reliability; conversely, a wide confidence interval suggests significant uncertainty, potentially requiring further data analysis or model optimization to improve accuracy. This method not only enhances the accuracy of power load forecasting but also provides a more reliable decision-making basis for smart grid operation and scheduling.

[0084] This invention employs deep learning technology for power load forecasting. First, historical power load data and external influencing factors (such as weather and holidays) are collected and normalized. This invention particularly emphasizes the integration of multi-dimensional data, laying the foundation for subsequent multi-feature fusion. Compared to existing technologies that use Long Short-Term Memory (LSTM) networks to process time-series data, this invention uses Convolutional Neural Networks (CNNs) to extract temporal features through convolutional and pooling layers. CNNs have advantages in capturing local patterns and spatial features, transforming time-series data into image-like structures for processing, thereby extracting richer latent patterns. During CNN training, the Monte Carlo method is introduced, quantifying the uncertainty of model parameters through multiple random dropouts of neurons. This strategy helps assess uncertainty in prediction, enhancing the model's robustness, especially in dealing with volatility and sudden events. The features extracted by the CNN are then input into a deep forest model for in-depth learning and prediction. The deep forest model, based on decision tree ensembles, can handle high-dimensional features and enhance the model's generalization ability, while combining external factors to achieve multi-dimensional information fusion, improving prediction accuracy. By performing multiple samplings using the Monte Carlo method and calculating the variance of the predicted values, an uncertainty assessment is provided, making the prediction results more valuable and facilitating more flexible scheduling of the power system.

[0085] Example 3:

[0086] This invention also discloses a deep learning-based power supply trend prediction model structure, such as... Figure 3 As shown, it includes:

[0087] This invention presents a smart grid power load accuracy prediction model, which is a hybrid architecture integrating convolutional neural networks (CNN), deep forests, and Monte Carlo uncertainty quantification mechanisms. The overall structure consists of multiple core modules, which work together to achieve accurate power load prediction. Specifically, these modules include:

[0088] The time-series data reconstruction and preprocessing module is the model's input layer. It transforms the raw data into a format suitable for CNN model processing, laying the foundation for subsequent feature extraction and prediction. Specifically:

[0089] First, historical power load data (15-minute granularity) and external factor data (temperature, humidity, holidays, etc.) are collected. Then, the one-dimensional time series is converted into a two-dimensional matrix by using a sliding window (time dimension) and multivariate channels (feature dimension, such as load, temperature, and humidity corresponding to different channels). Next, the Dynamic Time Warping (DTW) algorithm is used to filter historical similar day data to enhance the consistency of operating conditions of training samples. Finally, the data is scaled to a uniform range (e.g., [0,1]) to avoid the impact of differences in feature value ranges on model training.

[0090] The CNN deep feature extraction module, as the core of the model's feature extraction, uses convolution and pooling operations to uncover temporal patterns in electricity load data. Specifically:

[0091] A dual-channel convolutional layer is employed. Channel 1 uses a 3×3 convolutional kernel to focus on extracting local temporal texture features (such as short-term load fluctuation patterns). Channel 2 uses a 5×5 convolutional kernel to focus on capturing long-term trend features (such as daily and weekly load change patterns). A cross-layer feature fusion mechanism is adopted: low-level features (output of convolutional layer 1) and high-level features (output of convolutional layer 2) are concatenated through skip connections to retain multi-scale feature information. A hybrid pooling layer is used to combine max pooling and average pooling to reduce data dimensionality while retaining key features (max pooling) and suppressing noise interference (average pooling).

[0092] The Monte Carlo uncertainty quantization module, positioned between CNNs and deep forests, quantifies prediction uncertainty through random operations, improving the reliability of results. Specifically:

[0093] During the CNN model training phase, Dropout regularization (e.g., dropout rate of 0.5) is introduced into the fully connected layers of the CNN to randomly drop some neurons, forcing the model to learn redundant representations and avoid overfitting. During the prediction phase, T (e.g., 100) Monte Carlo samplings are performed, with different neurons randomly activated in each forward propagation to generate multiple predicted values. Based on the results of multiple samplings, uncertainty is calculated, and the predicted mean (μ) and variance (σ²) are calculated to generate a 95% confidence interval (e.g., [μ-1.96σ, μ+1.96σ]) to quantify the uncertainty.

[0094] The deep forest multi-granularity decision module, serving as the model's prediction output layer, fuses multi-dimensional features through ensemble learning to ultimately generate electricity load prediction results.

[0095] A multi-granularity scanning layer is used to perform sliding window sampling (window size 100 / 200 / 400) on the features output by the CNN to generate multi-scale feature fragments and enhance feature diversity. Then, a cascaded forest structure* is used. The first layer consists of a random forest and a fully random forest to select high-discrimination features. The second layer and above introduce an attention weighting mechanism to strengthen the decision weight of key features (such as the nonlinear relationship between temperature and load). Furthermore, an adaptive adjustment mechanism is adopted to dynamically adjust the number of forest layers (up to 10 layers) according to the prediction error to avoid overfitting and improve the model's generalization ability.

[0096] The prediction output module is the output layer of the model, responsible for integrating all features and generating the final prediction result. It includes: fusing the temporal features extracted by CNN with features from external factors (weather, holidays, etc.); performing ensemble learning through a hierarchical forest structure (up to 10 layers) to dynamically optimize the prediction result; and outputting the predicted power load value and the corresponding uncertainty indicators (mean, variance, confidence interval).

[0097] In the prediction model of this invention, the time-series data reconstruction module, which involves the collaborative relationships between various modules, inputs preprocessed two-dimensional data into a CNN model. The multi-scale features extracted by the CNN model, after Monte Carlo uncertainty quantification, are input into a deep forest along with external features. The deep forest then outputs the final prediction result and uncertainty assessment through multi-granularity decision-making. This architecture leverages the advantages of the CNN model in capturing time-series features, utilizes the ensemble learning capabilities of the deep forest to fuse multi-dimensional information, and solves the uncertainty problem through the Monte Carlo method, thus achieving accurate prediction of power load.

[0098] The present invention also provides an electronic device, Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention, such as... Figure 4 As shown, the electronic device may include a processor, a communications interface, memory, and a communication bus, wherein the processor, communications interface, and memory communicate with each other via the communication bus. The processor can invoke logical instructions from the memory, for example, to execute the following method:

[0099] S1. Collect historical power load time-series data and perform time-series data reconstruction and preprocessing;

[0100] S2. Deep feature extraction is performed on the processed time-series data based on the trained CNN model;

[0101] S3. Based on the extracted deep features, multi-granularity decision-making is performed using the trained deep forest to obtain power load forecasts.

[0102] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0103] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments, including, for example:

[0104] S1. Collect historical power load time-series data and perform time-series data reconstruction and preprocessing;

[0105] S2. Deep feature extraction is performed on the processed time-series data based on the trained CNN model;

[0106] S3. Based on the extracted deep features, multi-granularity decision-making is performed using the trained deep forest to obtain power load forecasts.

[0107] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0108] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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; and these 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 the present invention.

Claims

1. A deep learning-based method for predicting power supply trends, characterized in that, include: S1. Collect historical power load time-series data and perform time-series data reconstruction and preprocessing; S2. Deep feature extraction is performed on the processed time-series data based on the trained CNN model; S3. Based on the extracted deep features, multi-granularity decision-making is performed using the trained deep forest to obtain power load forecasts.

2. The method according to claim 1, characterized in that, The time-series data reconstruction and preprocessing performed in S1 include: Reconstruct historical power load time series data into two-dimensional gridded data; A similar working condition matching mechanism is constructed, and historical similar day data is filtered through the DTW algorithm.

3. The method according to claim 1, characterized in that, The CNN model in S2 is based on the Monte Carlo dropout regularization method to quantify the uncertainty of the CNN model.

4. The method according to claim 3, characterized in that, Monte Carlo quantization of CNN model uncertainty includes: During the training phase of CNN models, a dropout regularization mechanism is introduced into the fully connected layers to force the model to learn redundant representations. During the prediction phase of the CNN model, multiple Monte Carlo samplings are performed, with different neurons being randomly activated during each forward propagation.

5. The method according to claim 4, characterized in that, When performing multiple Monte Carlo samplings, the prediction results from these multiple samplings are used to statistically calculate the mean µ and variance. The uncertainty of CNN models is analyzed and quantified by the following formula: Where T is the number of Monte Carlo samples, y t These are the predicted values ​​from a single sample, where µ is the mean. Let Variance be the variance.

6. The method according to claim 1, characterized in that, The CNN model in S2 extracts deep features, including: Based on a dual-channel convolutional structure, local temporal texture features are extracted and long-term trend features are captured. Low-level features and high-level features are spliced ​​together by skip connections; The pooling layer employs hybrid max-average pooling, which preserves key features while suppressing noise.

7. The method according to claim 1, characterized in that, The deep forest multi-granularity decision-making process in S3 includes: Based on the multi-granularity scanning layer, sliding window sampling is performed on the deep features output by the CNN to generate multi-scale feature fragments; A combination of random forest and fully random forest is used to screen for high-discrimination features; An attention-weighted mechanism is introduced to strengthen the decision-making weight of key features.

8. The method according to claim 7, characterized in that, During deep forest training, an adaptive enhancement strategy is adopted to dynamically adjust the number of forest layers based on the prediction error.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 8.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 8.

Images