Modal classification and heterogeneous deep learning fusion-based power prediction method

Abstract

The invention discloses a modal classification and heterogeneous deep learning fusion-based power prediction method. The method comprises the steps of converting a timestamp corresponding to power time sequence data into a time feature; decomposing power data in the power time sequence data into N intrinsic mode function component sequences and a residual term component sequence; classifying each component sequence into one of four preset categories; matching corresponding heterogeneous deep learning sub-models for different types of component sequences; constructing a component data set according to the time characteristic and the power data of each component sequence, and dividing the component data set into a training set, a verification set and a test set; fitting an independent normalizer for the training set, and constructing a time window and a prediction step length suitable for supervised learning; independently training the heterogeneous deep learning sub-model according to the normalized training set; predicting a test set by using the trained heterogeneous deep learning sub-model; and carrying out reverse normalization on the prediction result of each heterogeneous deep learning sub-model through a respective normalizer, and finally adding all prediction values after reverse normalization by adopting a summation fusion strategy to obtain a final power prediction result. According to the method, the signal components are accurately classified, and a special model is used, so that the prediction precision and robustness are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of industrial data analysis and artificial intelligence, in particular to a power prediction method based on modal classification and heterogeneous deep learning fusion for industrial processes such as power load and energy consumption, and specifically relates to the integrated application of signal processing, machine learning and deep learning models. BACKGROUND

[0002] Industrial power load prediction is a core technical support for economic dispatching of power grids, optimization of energy allocation, and safety control of production, especially in energy-intensive enterprises such as coal mines. Accurate power prediction is directly related to the continuous operation guarantee of mining equipment, the stability of mine power supply systems, and the control of energy consumption costs, and has important practical significance for promoting the green and low-carbon transformation and safety production standardization of the coal industry.

[0003] However, the power operation data of coal mining enterprises presents typical multi-dimensional complexity characteristics, specifically: first, strong non-stationarity and trend drift, influenced by factors such as mine excavation plan adjustment, face alternation, and equipment maintenance cycle, power data often accompanies significant long-term trend and local mean shift, traditional stationarity assumption is completely invalid; second, multi-scale periodicity coupling, data is superimposed with multiple rhythms such as 24-hour day-night power supply cycle, 8-hour "three-shift" production cycle and 7-day weekly production plan cycle, each cycle component amplitude and phase interfere with each other; third, high-frequency nonlinear mutation, influenced by instantaneous working condition changes such as fully mechanized support start-stop, scraper conveyor load fluctuation, there are a large amount of high-frequency noise and nonlinear mutation burrs in the data, the signal-to-noise ratio is significantly low; fourth, multi-component strong coupling, the above trend, periodic and noise components do not exist independently, but present a deep interwoven coupling state, making it extremely difficult to extract features from the original signal.

[0004] In view of the above complex characteristics, traditional prediction methods have shown obvious limitations: statistical methods based on time series stationarity assumption (such as ARIMA model), its linear modeling framework cannot capture the strong non-linear relationship and multi-period coupling characteristics in power data, the prediction error is significant; a single deep learning model (such as LSTM, GRU) has certain non-linear fitting ability, but is limited by the unified network structure design, it is difficult to model the coupled multi-components differently - it is easy to fall into local optimum when dealing with high-frequency noise, and there is a problem of insufficient feature capture due to gradient disappearance when fitting trend items, ultimately presenting a "trade-off" prediction bottleneck.

[0005] To solve this problem, the "decomposition-integration" strategy has become the mainstream technical path in the field of complex time series prediction. Empirical mode decomposition (EMD) and its improved algorithm (such as CEEMDAN) can decompose the original power signal into a series of intrinsic mode functions (IMF) and residual trend items with a single characteristic scale, laying a foundation for subsequent accurate prediction. However, the existing decomposition-integration method generally adopts the "same modeling" paradigm, which means that the same structure of the prediction model is configured for all the IMF components and trend items. The core defect of this strategy is that it ignores the essential differences between different components. Different IMF components correspond to different frequency scale physical processes (such as high-frequency components corresponding to instantaneous disturbances of equipment, medium-frequency components corresponding to production shift cycles, and low-frequency components corresponding to long-term production trends). Their time correlation, fluctuation amplitude, and change rules are significantly different.

[0006] This "one-size-fits-all" modeling approach directly leads to double modeling bias. If a complex deep learning model (such as deep LSTM) is used to predict high-frequency noise components, not only is the computational resource wasted, but the model is also more likely to overfit due to the randomness of noise. If a simple linear model (such as ELM) is used to fit the low-frequency trend item, it is difficult to capture the nonlinear change characteristics of the trend item, resulting in underfitting bias. In essence, the existing method fails to establish a matching mechanism between component characteristics and model structure, which is the core technical bottleneck restricting the improvement of prediction accuracy. Therefore, it is urgent to build a heterogeneous modeling system based on component feature recognition to achieve accurate matching between different IMF components and optimal prediction models. SUMMARY

[0007] The present application aims to solve the problems of single model difficulty in fitting complex signals and low efficiency and accuracy caused by the use of homogeneous models in traditional decomposition-integration methods. The present application provides a power prediction method based on modal classification and heterogeneous deep learning fusion.

[0008] The technical solution of the present application is:

[0009] According to the first aspect of the present application, a power prediction method based on modal classification and heterogeneous deep learning fusion is provided, comprising:

[0010] S1, obtaining power time series data; converting the time stamp corresponding to the power time series data into a time feature; wherein the power time series data includes a time stamp and power data;

[0011] S2, decomposing the power data in the power time series data into N intrinsic mode function component sequences and a residual component sequence;

[0012] S3. For each component sequence obtained in S2, stationary analysis and spectral periodicity analysis are sequentially performed, and the component sequence is classified into one of four preset categories: high-frequency noise component, low-frequency noise component, seasonal component, or trend component;

[0013] S4. According to the classification result of S3, a corresponding heterogeneous deep learning sub-model is matched for the component sequence of different categories;

[0014] S5. A component data set is constructed according to the time characteristics and power data of each component sequence, and is divided into a training set, a validation set, and a test set. An independent normalizer is fitted for the training set, and a time window and a prediction step suitable for supervised learning are constructed;

[0015] S6. According to the normalized training set, the N+1 heterogeneous deep learning sub-models constructed in S4 are independently trained, and the validation set and early stopping mechanism are used to prevent overfitting;

[0016] S7. The trained heterogeneous deep learning sub-models are used to predict the test set. The prediction results of each heterogeneous deep learning sub-model are de-normalized through the respective normalizer, and finally a summation fusion strategy is used to add all the de-normalized prediction values to obtain the final power prediction result.

[0017] Further, the time characteristics include hour, week, and month characteristics based on sine and cosine transformation.

[0018] Further, S2 specifically comprises: using a completely adaptive noise ensemble empirical mode decomposition algorithm to decompose the power data in the power time series data into N intrinsic mode function component sequences and a residual component sequence.

[0019] Further, S3 specifically comprises:

[0020] S3.1. ADF test is performed on the first component sequence to obtain a p value ; a stationary threshold is set, and if < , the component sequence is determined to be stationary; if , the component sequence is determined to be non-stationary; wherein ;

[0021] S3.2. If the component is determined to be non-stationary in S3.1, or the component is the residual component sequence obtained by decomposition, the component is classified as a trend component; if the component For stationary, FFT analysis is performed on it, its normalized amplitude spectrum is calculated, and a peak detection algorithm is used to find all spectral peaks whose height exceeds a peak significance threshold;

[0022] S3.3, for all spectral peaks detected in S3.2, check whether their corresponding frequency values match the preset key business cycle frequency: if the frequency value of any spectral peak detected falls within the adjacent range of any key business cycle frequency, it will be classified as a seasonal component;

[0023] S3.4, if S3.2 determines that is stationary, and no spectral peak matching the key business cycle frequency is found, the component sequence is determined to be noise; in the case of determining noise, the dominant frequency of the component sequence is calculated , and compared with the high-frequency noise threshold : if , , it is classified as a high-frequency noise component; if , it is classified as a low-frequency noise component.

[0024] Further, the different categories of components are matched with corresponding heterogeneous deep learning sub-models, as follows:

[0025] S4.1, for components of the category of high-frequency noise, and residual terms of the category of trend (i=N+1), a lightweight LSTM model is assigned;

[0026] S4.2, for components of the category of low-frequency noise, and non-residual terms of the category of trend, a standard three-layer LSTM model is assigned;

[0027] S4.3, for components of the category of seasonality, a seasonal frequency model is assigned.

[0028] Further, the seasonal frequency model includes:

[0029] The input is linearly mapped to a high-dimensional time-domain representation ;

[0030] The time-domain representation is subjected to a fast Fourier transform to obtain ; the real and imaginary parts of the complex number are processed using an MLP to obtain ; the processed frequency-domain information is converted back to the time domain ;

[0031] The original embedding is linearly transformed as a residual connection to obtain ; the processing result of converting the frequency domain to the time domain is added to the time-domain residual , and fed into an MLP to obtain ;

[0032] Time-domain MLP processing results The final output is a prediction vector through layer normalization, flattening, and a fully connected prediction head.

[0033] Furthermore, the sub-model training process of S6 is implemented as follows:

[0034] S6.1 Setting the loss function: The mean squared error is used as the loss function for all sub-models;

[0035] S6.2 Setting the optimizer: The Adam optimizer is used in conjunction with a learning rate decay strategy to automatically adjust the learning rate based on the validation set loss;

[0036] S6.3 Early Stop: After each training cycle, evaluate the model loss on the validation set. If the validation set loss does not decrease further within the preset tolerance value, terminate the training early and load the saved best model weights.

[0037] According to a second aspect of the present invention, a power prediction system based on the fusion of modality classification and heterogeneous deep learning is provided, comprising modules of any of the methods described above.

[0038] According to a third aspect of the present invention, a processor is provided for running a program, wherein the program, when running, performs the steps of the lithium-ion battery RUL prediction method based on sequence decomposition and the PatchTST-MKAN model as described above.

[0039] According to a fourth aspect of the present invention, a terminal device is provided, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of any of the methods described above.

[0040] The beneficial effects of this invention are:

[0041] This invention abandons the inefficient strategy of "homogeneous models" and uses "stationarity analysis and spectral periodicity analysis" for scientific classification. It matches specially optimized "heterogeneous models" to signal components with different characteristics, ensuring optimal fitting for each component. The final summation and fusion accuracy is significantly higher than that of a single model or a homogeneous ensemble model. The classification step provides a physical explanation for the "black box" decomposition-ensemble process. Through FFT analysis, it is clear which components correspond to "three shifts" (8-hour cycle) and which to "day and night" (24-hour cycle), making model decisions more transparent. This invention significantly improves prediction accuracy and robustness by accurately classifying signal components and using specialized models (for simple components such as high-frequency noise, this invention uses a lightweight model (LSTM2), greatly reducing the total number of model parameters and training time, while avoiding the risk of overfitting). Attached Figure Description

[0042] Figure 1 This is a flowchart of the present invention.

[0043] Figure 2 The waveforms of each IMF component and the residual obtained after CEEMDAN decomposition of the original power time series data are shown.

[0044] Figure 3 This is a graph showing the predicted and actual values ​​of the method of this invention on the test set. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0046] Example 1: As Figures 1-3 As shown, according to a first aspect of the present invention, a power prediction method based on the fusion of modality classification and heterogeneous deep learning is provided, comprising:

[0047] S1. Obtain power time series data; convert the timestamp column corresponding to the power time series data into time features, the time features including hour, weekday, and month features based on sine and cosine transforms; wherein, the power time series data includes a time timestamp column and a power data column;

[0048] S2. The fully adaptive noise ensemble empirical mode decomposition (CEEMDAN) algorithm is used to decompose the power data column in the power time series data into N intrinsic mode function (IMF) component sequences and one residual term (R) component sequence, collectively referred to as N+1 component sequences.

[0049] S3. For each component sequence obtained in S2, perform stationarity analysis and spectral periodicity analysis in sequence, and classify it into one of the four preset categories: high frequency noise component (HF_NOISE), low frequency noise component (LF_NOISE), seasonal component (SEASONAL), or trend component (TREND).

[0050] S4. Based on the classification results of S3, match the corresponding heterogeneous deep learning sub-models for the component sequences of different categories.

[0051] S5. Based on the classification results of S3, construct component datasets according to the time features and power data of each component sequence, and divide them into training set, validation set and test set; fit independent normalizers (such as MinMaxScaler) to the training set, and construct time windows (look-back) and prediction step sizes (horizon) suitable for supervised learning.

[0052] S6. Based on the normalized training set, independently train the N+1 heterogeneous deep learning sub-models constructed in S4, and use a validation set and early stopping mechanism to prevent overfitting.

[0053] S7. Use the trained heterogeneous deep learning sub-models to predict the test set; denormalize the prediction results of each heterogeneous deep learning sub-model through its own normalizer, and finally use a summation and fusion strategy to add up all the denormalized prediction values ​​to obtain the final power prediction result.

[0054] Furthermore, S3 specifically refers to:

[0055] S3.1, regarding the first Component sequence Perform the ADF (Augmented Dickey-Fuller) test to obtain the p-value. Set a stability threshold. (For example =0.05), if < If the component sequence is stationary, then it is determined that the component sequence is stationary; if If the component sequence is non-stationary, then it is determined that the component sequence is non-stationary; where, ;

[0056] S3.2, If the determination in S3.1 is correct... It is non-stationary, or the component The final residual components of the fully adaptive noise set empirical mode decomposition (i.e. If the corresponding residual component is used, it is classified as a trend component; if the determination in S3.1 is correct... To ensure stability, an FFT (Fast Fourier Transform) analysis is performed on it to calculate its normalized amplitude spectrum, and a peak detection algorithm is used to find all spectral peaks whose height exceeds the peak significance threshold (e.g., 0.3).

[0057] S3.3 For all spectral peaks detected in S3.2, check whether their corresponding frequency values ​​match the preset key business cycle frequency: If the frequency value of any detected spectral peak falls within the vicinity of any key business cycle frequency (e.g., allow a deviation of ±0.05), then the component sequence is determined to have significant periodic characteristics and is classified as a seasonal component; whereby the key business cycle frequency is preset based on the actual business operation rules, including: daily cycle (f=1.0, i.e., a cycle once every 24 hours), half-day cycle (f=2.0, i.e., a cycle once every 12 hours), eight-hour cycle (f=3.0, i.e., a cycle once every 8 hours, corresponding to a three-shift work system) and weekly cycle (f=1 / 7, i.e., a cycle once every 7 days).

[0058] S3.4, If the determination in S3.2 is correct... If the signal is stable and no spectral peaks matching the frequency of the critical business cycle are found, the component sequence is considered noise. In the case of noise, the dominant frequency of the component sequence is calculated. (Frequency at maximum amplitude), and compared with the high-frequency noise threshold. Comparison: If > If it is classified as a high-frequency noise component; if If so, it is classified as a low-frequency noise component; for example .

[0059] Furthermore, the heterogeneous deep learning sub-models corresponding to component matching of different categories are as follows:

[0060] S4.1 For the component classified as high-frequency noise and the residual term classified as trend (i=N+1), a lightweight LSTM model is assigned. The lightweight LSTM model includes two stacked "LSTM layers + Dropout layers", a fully connected layer, and a linear layer. The model uses two LSTM layers (LSTM2Layer_Small) to have fewer hidden layers (e.g., 2 layers) and fewer hidden units (e.g., [128, 64]).

[0061] S4.2 For components classified as low-frequency noise and non-residual terms classified as trends, a standard three-layer LSTM model (LSTM3Layer) is assigned (each layer consists of "LSTM layer + Dropout layer", for a total of three layers). This model has deeper hidden layers and more hidden units than the lightweight LSTM model in order to capture complex low-frequency and trend dynamics.

[0062] S4.3 For components classified as seasonal, assign a seasonal frequency model (SFM) specifically designed for handling strongly periodic signals.

[0063] Furthermore, the seasonal frequency model (SFM) includes a forward propagation process, which comprises:

[0064] S5.1, Temporal Embedding: ,Will Input By linear mapping to a high-dimensional time-domain representation ;

[0065] S5.2. Perform a Fast Fourier Transform on the time-domain representation, the expression is: Use a frequency domain MLP to process complex numbers separately. The real and imaginary parts are expressed as follows: The processed frequency domain information is converted back to the time domain using the inverse Fourier transform (IFFT). , ;

[0066] S5.3, Perform a linear transformation on the original embedding as a residual connection, i.e. The result of converting the frequency domain to the time domain is added to the time domain residual and then fed into the time-domain MLP (Mechanical Processing Level). ), ;

[0067] S5.4. The temporal MLP processing results are processed through layer normalization (LayerNorm), flattening, and a fully connected prediction head (FC_Head) to finally output an H-dimensional prediction vector. .

[0068] Furthermore, the input of the model defined in S5 based on the component dataset... and output as follows:

[0069] ;

[0070] ;

[0071] In the formula, For one The characteristic matrix is ​​composed of normalized component values ​​at L time steps. and K time features at L time steps constitute; Let be an H-dimensional prediction vector, representing the normalized component values ​​for the next H time steps.

[0072] Furthermore, the model training process is as follows:

[0073] S6.1 Setting the loss function: All sub-models use mean squared error (MSE) as the loss function. The calculation formula is as follows:

[0074] ;

[0075] In the formula, M is the batch size, and H is the prediction step size. For the true value, This is a predicted value;

[0076] S6.2 Setting the optimizer: Use the Adam optimizer and a learning rate decay strategy (such as ReduceLROnPlateau) to automatically adjust the learning rate based on the validation set loss;

[0077] S6.3 Early Stop: After each training epoch, evaluate the model loss on the validation set. If the validation set loss does not decrease further within a preset patience value (e.g., 20 epochs), terminate training early and load the saved optimal model weights.

[0078] According to a second aspect of the present invention, a power prediction system based on the fusion of modality classification and heterogeneous deep learning is provided, comprising modules of any one of the methods described above, specifically including: an acquisition module, used to acquire power time series data; convert the timestamps corresponding to the power time series data into time features; a CEEMDAN decomposition module, used to decompose the power data in the power time series data into N intrinsic mode function component sequences and a residual term component sequence using a fully adaptive noise ensemble empirical mode decomposition algorithm; a modality classification module, used to perform stationarity analysis and spectral periodicity analysis on each component sequence in sequence, classifying it into one of four preset categories: high-frequency noise component, low-frequency noise component, seasonal component, or trend component; and a matching module, used to match different... The system comprises the following modules: a component sequence matching module for each category and a corresponding heterogeneous deep learning sub-model; a partitioning module to construct a component dataset based on the temporal features and power data of each component sequence, dividing it into training, validation, and test sets; fitting independent normalizers to the training set and constructing time windows and prediction step sizes suitable for supervised learning; a training module to independently train N+1 heterogeneous deep learning sub-models based on the normalized training set, employing a validation set and early stopping mechanism to prevent overfitting; a prediction module to predict the test set using the trained heterogeneous deep learning sub-models; and a summation and fusion module to denormalize the prediction results of each heterogeneous deep learning sub-model using its respective normalizer, and finally summing all the denormalized prediction values ​​using a summation and fusion strategy to obtain the final power prediction result. Each module in the above power prediction system based on the fusion of modality classification and heterogeneous deep learning can be implemented entirely or partially through software, hardware, or a combination thereof. The above modules can be embedded in the processor of the computer device in hardware form or independent of it, or they can be stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of the above modules.

[0079] According to a third aspect of the present invention, a processor is provided, the processor being configured to run a program, wherein the program, when running, performs the steps of the lithium-ion battery RUL prediction method based on sequence decomposition and the PatchTST-MKAN model as described above.

[0080] According to a fourth aspect of the present invention, a terminal device is provided, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of any of the methods described above.

[0081] Example 2: The method of the present invention is explained below with reference to the collected power consumption data of Rong'an Coal Mine:

[0082] Step 1: Collect power time series data In this embodiment, the power time series data is power data collected from Rong'an Coal Mine at 15-minute sampling intervals. Each power time series data includes time and power consumption. The timestamp column in the collected power time series data is converted into K time features. To solve the problem of periodic jumps in time features (such as hours, days of the week) (e.g., from 11 PM to midnight), this invention uses sine and cosine transforms for encoding. The K time features are specifically:

[0083] ;

[0084] ;

[0085] ;

[0086] ;

[0087] ;

[0088] ;

[0089] in, , These represent the hourly characteristics based on sine and cosine transforms, respectively. , These represent the weekday characteristics based on sine and cosine transforms, respectively. , These represent the month characteristics based on sine and cosine transforms, respectively. This indicates the number of hours corresponding to the time period in the power time series data. This indicates the week number corresponding to the time in the power time series data. This indicates the month number corresponding to the time in the power time series data (e.g., 15:18 on December 16, 2025). Take 15. Take 2, Take 12).

[0090] Step 2: Employ the Fully Adaptive Noise Ensemble Empirical Mode Decomposition (CEEMDAN) algorithm to decompose the power data series into N intrinsic mode function (IMF) components and one residual (R) component, collectively referred to as the N+1 component sequence, as shown in the following expression:

[0091] ;

[0092] in, Representing the Component sequence For residual components ( In this embodiment, trials=50 is set, resulting in N+1 = 16 components. The decomposition results are as follows. Figure 2 As shown, the component sequences from top to bottom are IMF 1 → IMF 15, Residue (i.e. The frequencies of the IMF components decrease sequentially from top to bottom. Figure 2 As can be intuitively seen, the original power signal, which originally contained complex nonlinear and non-stationary characteristics, was effectively decomposed into several intrinsic mode function (IMF) components with different time scales and a final trend residual term. Among them, the first IMF components (such as IMF1 and IMF2) have higher frequencies and more dramatic fluctuations, reflecting short-term rapid changes and high-frequency noise in the signal; as the order increases, the frequencies of the IMF components gradually decrease, and the waveforms tend to be smoother, reflecting medium- and long-term periodic fluctuations in the signal; the final residual term represents the overall long-term evolution trend of the power load. Figure 2 The decomposition shown decouples the complex mixed signal into single components with different characteristics, laying the foundation for accurate modal feature analysis and classification of different components in the subsequent third step.

[0093] Step 3: Process the N+1 components obtained in step 2 Each component is scientifically "diagnosed" and classified into high-frequency noise components (HF_NOISE), low-frequency noise components (LF_NOISE), seasonal components (SEASONAL), or trend components (TREND).

[0094] 1) For component sequences Perform the ADF (Augmented Dickey-Fuller) test to obtain the p-value. Set a stability threshold. = 0.05. If Then determine It is a stationary sequence. If... Then determine It is a non-stationary sequence.

[0095] 2) If the judgment in 1) If it is non-stationary, then the component has a trend. If the component... The final residual term R(t) of CEEMDAN (i.e. If any of the above conditions are met, the component will be classified as TREND.

[0096] 3) If the judgment in 1) is correct To ensure stability, an FFT (Fast Fourier Transform) analysis is performed to calculate its normalized amplitude spectrum and to find all spectral peaks whose height exceeds the peak significance threshold (set to 0.3 in this embodiment).

[0097] 4) Check whether the spectral peaks detected in 3) match the preset key business cycle frequency: If the frequency value of any detected spectral peak falls within the vicinity of any key business cycle frequency (for example, a deviation of ±0.05 is allowed), then the component sequence is determined to have significant periodicity characteristics and is classified as SEASONAL; where the key business cycle frequency is preset based on the actual business operation rules, including: daily cycle (f=1.0, i.e., a cycle once every 24 hours), half-day cycle (f=2.0, i.e., a cycle once every 12 hours), eight-hour cycle (f=3.0, i.e., a cycle once every 8 hours, corresponding to a three-shift work system) and weekly cycle (f=1 / 7, i.e., a cycle once every 7 days).

[0098] 5) If If a component is determined to be stationary and not classified as SEASONAL in 4), then that component is noise. Calculate its dominant frequency. (Frequency at maximum amplitude), and compared with the high-frequency noise threshold. (This embodiment is set as follows) This value is slightly higher than the 8-hour period (f=3.0) for comparison: If If so, it is classified as HF_NOISE (high-frequency noise). If so, it is classified as LF_NOISE (low-frequency noise / drift).

[0099] Step 4: Based on the classification results of Step 3, classify the component sequences into different categories. Assigning deep learning sub-models with different structures : Assign an LSTM2Layer_Small model (a lightweight 2-layer LSTM with hidden units [128, 64]) to the residual terms (IMF16 in this example) in the HF_NOISE and TREND categories. Assign an LSTM3Layer model (a standard 3-layer LSTM with hidden units [256, 128, 64]) to the non-residual terms in the LF_NOISE and TREND categories. Assign a (SFM) model (a dedicated frequency model combining FFT and MLP) to the SEASONAL category.

[0100] Step 5: Data Normalization and Dataset Construction

[0101] 1) Construct component datasets based on the temporal characteristics and power data of each component sequence. The datasets are divided into training set (70%), validation set (15%), and test set (15%) in chronological order.

[0102] 2) Due to the significant differences in the amplitude (dimensions) of each IMF component (e.g.) Figure 2 As shown), a separate normalizer must be used. For each of the N+1 components, instantiate N+1 MinMaxScaler normalizers. Using training set data to... Perform a fit, and then transform the training set, validation set, and test set respectively;

[0103] 3) Input and output of the model defined using the sliding window method: Set the time window L and prediction step size H. In this embodiment: L=96, H=24 for each component sequence. Construct its input-output pair At time t: Input :one The feature matrix contains and [ T t−L+1 ,⋯T t ] Output A prediction vector of dimension H .

[0104] Step 6: Feed the N+1 training and validation sets constructed in Step 5 into the corresponding N+1 heterogeneous deep learning sub-models from Step 4. Independent training is conducted during this period.

[0105] 1) All sub-models use mean squared error (MSE) as the loss function. The calculation formula is as follows:

[0106] ;

[0107] Where M is the batch size and H is the prediction step size. For the j-th sample, the true value of the k-th step size. This is the corresponding predicted value.

[0108] 2) The Adam optimizer is used in conjunction with the ReduceLROnPlateau learning rate decay strategy. During training, at the end of each epoch, the learning rate is evaluated on the validation set. If the validation set loss does not decrease further within the preset patience value (20 in this embodiment), training is terminated early, and the saved optimal model weights are loaded.

[0109] Step 7: Prediction and Integration.

[0110] 1) Test set Feed each of the N+1 pre-trained sub-models Normalized predicted values ​​are obtained. .

[0111] 2) Use the N+1 normalizers saved in step 5 ,right Perform inverse normalization to obtain component predictions with true physical scale. .

[0112] 3) Based on the CEEMDAN decomposition principle in step two The predicted values ​​of the N+1 inversely normalized components The elements are added one by one at the corresponding time steps to obtain the final power prediction result. :

[0113] ;

[0114] The embodiments of this invention use root mean square error. and coefficient of determination The following are some of the metrics used to evaluate the predictive performance of ensemble models:

[0115] ;

[0116] ;

[0117] In the formula, the root mean square error (RMSE) and the coefficient of determination (R²) are used as evaluation indicators for the model's predictive performance, M is the total number of test samples, and H is the prediction step size. For the true value, For predicted values, The mean of the true values. The smaller the value and The larger the value, the better the model's predictive performance.

[0118] The prediction performance of this invention was compared with that of traditional LSTM and single-mode decomposition methods on coal mine power datasets (LSTM directly uses a two-layer LSTM model to predict the original power time series data, while the single-mode decomposition method uses the same fully adaptive noise set empirical mode decomposition algorithm as this invention to decompose the data and then uses LSTM to predict each component). The results are shown in Table 1.

[0119] Table 1 Comparison of prediction performance of different methods on the coal mine power dataset

[0120]

[0121] As shown in Table 1, the modeling method proposed in this invention exhibits excellent predictive performance on the test set. Its coefficient of determination (R²) reaches 0.9127, indicating that the method can highly fit the changing trends of the real data and explain more than 91% of the data variance. This demonstrates that the present invention, through its strategies of "scientific classification" and "heterogeneous modeling," effectively distinguishes the characteristics of different signals and uses the most suitable model for fitting, exhibiting a significant accuracy advantage compared to single models or homogeneous ensemble models.

[0122] To verify the effectiveness of the power prediction method based on the fusion of modality classification and heterogeneous deep learning proposed in this invention, a verification experiment was conducted using the partitioned independent test set data. Figure 3 The diagram shows a comparison curve between the predicted power values ​​and the actual power values ​​for a specific time period on the test set using the method of this invention. Figure 3 In the diagram, the solid curve represents the actual power load data, and the dashed curve represents the final predicted output of the fusion model of this invention. By observing... Figure 3 It is clearly evident that the predicted curve obtained by the method of this invention has an extremely high degree of fit with the actual value curve. The model not only accurately captures the main periodic trends and long-term trends of the power load data, but also closely follows the changes in the actual value at key turning points such as peaks and troughs, and at moments of rapid load fluctuation. This high degree of consistency demonstrates that by decomposing the complex power signal into different modes and specifically utilizing lightweight LSTM models, standard three-layer LSTM models, and SFM models for separate modeling and fusion, the model's ability to perceive and predict power load under complex operating conditions is effectively improved, proving that the method of this invention possesses excellent predictive performance and good generalization ability.

[0123] The above embodiments are used to explain the effectiveness and feasibility of the present invention. The key point of this invention is to decouple highly coupled complex time series into multiple components using signal decomposition, and to innovatively introduce modal classification (ADF+FFT) to automatically identify the time-frequency characteristics of each component (such as high-frequency noise, low-frequency noise, specific periodicity, and trends). This allows for matching optimal heterogeneous deep learning sub-models to different components, and finally, high-precision fusion is achieved through summation. The introduction of this complete "decomposition-classification-heterogeneity-fusion" strategy aims to address the problems of strong non-stationarity, multiple periodicity, strong coupling of high-frequency noise, and poor predictive performance of conventional single models or homogeneous ensemble models in industrial process data.

[0124] It should be noted that this invention is not only applicable to coal mine power load forecasting, but can also be widely applied to various industrial and commercial energy consumption scenarios with periodic and multi-scale characteristics. For example, in steel smelting, it can predict the impact load of electric arc furnaces and the periodic power fluctuations of rolling mills; in chemical plants, it can accurately analyze the DC power of electrolytic cells and the heat energy consumption changes of reactors; in commercial building complexes, it can optimize the daily cycle energy consumption of central air conditioning and lighting systems; in data centers, it can predict server power consumption and the dynamic load of cooling systems; in the rail transit field, it is suitable for peak and valley power analysis of subway traction power supply; and in new energy power stations, it can combine weather factors to predict the intermittent output of photovoltaic and wind power.

[0125] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A power prediction method based on the fusion of modality classification and heterogeneous deep learning, characterized in that, include: S1. Obtain power time series data; convert the timestamps corresponding to the power time series data into time features; wherein, the power time series data includes timestamps and power data; S2. Decompose the power data in the power time series data into N intrinsic mode function component sequences and one residual term component sequence; S3. For each component sequence obtained in S2, perform stationarity analysis and spectral periodicity analysis in sequence, and classify it into one of the four preset categories: high-frequency noise component, low-frequency noise component, seasonal component, or trend component. S4. Based on the classification results of S3, match the corresponding heterogeneous deep learning sub-models for the component sequences of different categories. S5. Construct a component dataset based on the time features and power data of each component sequence, and divide it into a training set, a validation set, and a test set; fit an independent normalizer to the training set, and construct a time window and prediction step size suitable for supervised learning. S6. Based on the normalized training set, independently train the N+1 heterogeneous deep learning sub-models constructed in S4, and use a validation set and early stopping mechanism to prevent overfitting. S7. Use the trained heterogeneous deep learning sub-models to predict the test set; denormalize the prediction results of each heterogeneous deep learning sub-model through its own normalizer, and finally use a summation and fusion strategy to add up all the denormalized prediction values ​​to obtain the final power prediction result.

2. The power prediction method based on the fusion of modality classification and heterogeneous deep learning according to claim 1, characterized in that, The time features include hour, weekday, and month features based on sine and cosine transforms.

3. The power prediction method based on the fusion of modality classification and heterogeneous deep learning according to claim 1, characterized in that, Specifically, S2 involves using a fully adaptive noise set empirical mode decomposition algorithm to decompose the power data in the power time series data into N intrinsic mode function component sequences and one residual term component sequence.

4. The power prediction method based on the fusion of modality classification and heterogeneous deep learning according to claim 1, characterized in that, Specifically, S3 is: S3.1, regarding the first Component sequence Perform an ADF test to obtain the p-value. Set a stability threshold. ,like If the component sequence is stationary, then it is determined that the component sequence is stationary; if If the component sequence is non-stationary, then it is determined that the component sequence is non-stationary; where, ; S3.2, If the determination in S3.1 is correct... It is non-stationary, or the component sequence If the residual component sequence is a series of decomposed terms, then it is classified as a trend component; if the determination in S3.1 is... To ensure stability, an FFT analysis is performed on it to calculate its normalized amplitude spectrum, and a peak detection algorithm is used to find all spectral peaks whose height exceeds the peak significance threshold. S3.3 For all the spectral peaks detected in S3.2, check whether their corresponding frequency values ​​match the preset key business cycle frequency: if the frequency value of any detected spectral peak falls within the vicinity of any key business cycle frequency, then classify it as a seasonal component. S3.4, If the determination in S3.2 is correct... If the signal is stable and no spectral peaks matching the frequency of the critical business cycle are found, the component sequence is considered noise. In the case of noise, the dominant frequency of the component sequence is calculated. and with high-frequency noise threshold Comparison: If If it is classified as a high-frequency noise component; if If it is, then it is classified as a low-frequency noise component.

5. The power prediction method based on the fusion of modality classification and heterogeneous deep learning according to claim 1, characterized in that, The above describes the heterogeneous deep learning sub-models corresponding to component matching of different categories, as follows: S4.

1. Assign a lightweight LSTM model to the components classified as high-frequency noise and the residuals classified as trends. S4.

2. Assign a standard three-layer LSTM model to the components classified as low-frequency noise and the non-residual terms classified as trends. S4.

3. Assign a seasonal frequency model to the component whose category is seasonal.

6. The power prediction method based on the fusion of modality classification and heterogeneous deep learning according to claim 5, characterized in that, The seasonal frequency model includes: The input is linearly mapped to a high-dimensional time-domain representation. ; Time-domain representation Perform a fast Fourier transform to obtain Using MLP to process the real and imaginary parts of complex numbers separately, we obtain... The processed frequency domain information is converted back to the time domain using the inverse Fourier transform. ; Perform a linear transformation on the original embedding as a residual connection to obtain The result of converting the frequency domain to the time domain. With time domain residual Add them together and feed them into the MLP to obtain... ; MLP processing results The final output is a prediction vector through layer normalization, flattening, and a fully connected prediction head.

7. The power prediction method based on the fusion of modality classification and heterogeneous deep learning according to claim 1, characterized in that, The implementation process of training the sub-model of S6 is as follows: S6.1 Setting the loss function: The mean squared error is used as the loss function for all sub-models; S6.2 Setting the optimizer: The Adam optimizer is used in conjunction with a learning rate decay strategy to automatically adjust the learning rate based on the validation set loss; S6.3 Early Stop: After each training cycle, evaluate the model loss on the validation set. If the validation set loss does not decrease further within the preset tolerance value, terminate the training early and load the saved best model weights.

8. A power prediction system based on the fusion of modality classification and heterogeneous deep learning, characterized in that, The module includes the method described in any one of claims 1-7.

9. A processor, characterized in that, The processor is used to run a program, wherein the program executes the steps of the lithium-ion battery RUL prediction method based on sequence decomposition and the PatchTST-MKAN model as described in any one of claims 1-7.

10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-7.

Images