Coal mine load prediction method based on adaptive multi-scale ema and multi-path signal feature fusion
By using an adaptive multi-scale EMA and multi-path signal feature fusion method, multi-dimensional input features are constructed and combined with an adaptive multi-scale EMA fusion module, solving the problem of unified characterization of trends and disturbances in coal mine load forecasting, and achieving high-precision and stable forecasting under complex working conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- KUNMING UNIV OF SCI & TECH
- Filing Date
- 2026-02-02
- Publication Date
- 2026-06-12
AI Technical Summary
Existing technologies are insufficient to effectively and uniformly characterize long-term trends, periodicity, and short-term disturbances in coal mine load forecasting, resulting in insufficient forecast accuracy and robustness, especially when the model performance is unstable during operating condition switching.
An adaptive multi-scale EMA and multi-path signal feature fusion method is adopted. Multi-dimensional input features are constructed through wavelet decomposition, K-Means clustering, time coding and two-layer LSTM. Combined with the adaptive multi-scale EMA fusion module, a dynamic adaptive balance between trend and disturbance is achieved.
Under complex working conditions in coal mines, it achieves high accuracy and stability in load forecasting, improving the accuracy and robustness of forecasts, especially with significant improvements in forecasting performance during working condition switching and short-term disturbance phases.
Smart Images
Figure CN122198208A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a coal mine load prediction method based on adaptive multi-scale EMA and multi-path signal feature fusion, belonging to the fields of artificial intelligence, deep learning and industrial intelligence. Background Technology
[0002] Coal mine safety is a core concern in the industrial sector. Electricity load sequences, as key time-series signals reflecting the operational status of coal mines, reveal underlying patterns in equipment conditions and production rhythms through their trend and periodic changes. Accurate load sequence forecasting is a prerequisite for timely detection of abnormal electricity consumption and the uncovering of potential production violations; therefore, load forecasting plays a crucial role in ensuring safe coal mine production.
[0003] However, achieving high-precision prediction of coal mine load remains a significant challenge. On one hand, the long-term trend, periodicity, and short-term disturbances coexist and mix in coal mine load sequences, resulting in highly volatile data, typical of non-stationary time series. Traditional statistical models (such as ARIMA) or general deep learning models relying solely on a single input path (such as standard LSTM) struggle to provide a unified and stable characterization of this multi-component time series dynamic. On the other hand, during the alternating trend and disturbance phases caused by shifts in operating conditions, load signals are prone to abrupt changes, and existing models lack effective mechanisms to balance stability during trend phases with agility during disturbance phases. For example, commonly used deep models or decomposition models often perform well in trend characterization but exhibit sluggish responses to short-term disturbances or excessive smoothing in extreme value phases, making it difficult to maintain prediction accuracy and robustness. Therefore, there is an urgent need to construct a unified prediction framework for complex coal mine operating conditions, capable of systematically processing the complex characteristics of input data and balancing long-term trends and short-term disturbances at both the modeling and output ends, thereby achieving both prediction accuracy and stability under complex operating conditions.
[0004] In view of this, the present invention is hereby proposed. Summary of the Invention
[0005] This invention provides a coal mine load forecasting method based on adaptive multi-scale EMA and multi-path signal feature fusion, for predicting highly volatile time-series load data.
[0006] The technical solution of this invention is:
[0007] According to a first aspect of the present invention, a coal mine load prediction method based on adaptive multi-scale EMA and multi-path signal feature fusion is provided, comprising:
[0008] S1. Obtain the raw time series data of coal mine electricity load and preprocess it to obtain the preprocessed time series data of electricity load.
[0009] S2. Sample window segmentation: The preprocessed electricity load time series data is segmented using a sliding window algorithm to obtain sample window data;
[0010] S3. Obtain wavelet decomposition features based on the sample window data. Clustering pattern characteristics Time coding features and smooth features ;
[0011] S4. Consolidate the wavelet decomposition features, clustering pattern features, time-coding features, and smoothing features along the feature dimension to generate multi-dimensional input features. ;in, ;
[0012] S5. Construct a double-layer stacked long short-term memory network; incorporate multi-dimensional input features. After normalization, it is used as the input of a stacked long short-term memory network. The hidden state output by the second LSTM is mapped step-by-step through a fully connected layer to obtain the original prediction sequence.
[0013] S6. Construct an adaptive multi-scale EMA fusion module; use the original prediction sequence as input to the adaptive multi-scale EMA fusion module and output the final prediction result.
[0014] Further, S3 includes:
[0015] S3.1 Decompose the sample window data using discrete wavelet transform to obtain wavelet decomposition features. ={ , , , The decomposition follows the following relationship:
[0016] ;
[0017] in, It is sample window data. It is the first Approximate components of the layer, It is the first Layer detail components;
[0018] set up Based on the above formula, we obtain four components: , , , ,in Characterizing the long-term trends and process rhythms of coal mine production, while , , This characterizes equipment start-up and shutdown, sudden fluctuations, and short-term disturbances with high-frequency spikes at different scales;
[0019] S3.2, Wavelet decomposition features KMeans clustering was performed to obtain cluster labels that reflect the operating patterns at each time point. Clustering labels pass The encoding is converted into sparse vectors, and the sparse vectors are stacked and combined into a matrix in chronological order to form clustering pattern features. ;
[0020] S3.3, Wavelet decomposition features using sine and cosine functions The time and location information is processed to construct intraday and intraweek periodic time codes, forming intraday features. and intra-week characteristics By concatenating intraday and intraweek features, time-coded features are obtained. ;
[0021] S3.4, Wavelet Decomposition Features Multi-scale exponential moving average smoothing is performed to obtain smoothing features. .
[0022] Furthermore, in the KMeans algorithm, the optimal number of clusters is determined using the elbow rule. .
[0023] Furthermore, cluster labels pass The process of converting the encoding into a sparse vector is as follows: Integer cluster labels at each time point are generated. Convert to a binary vector of length K, which is only valid at index equal to... The value in position 1 is 1, and the values in all other positions are 0.
[0024] Furthermore, in S5, a Dropout layer with a dropout rate of 0.2 is introduced after each LSTM layer.
[0025] Furthermore, the adaptive multi-scale EMA fusion module is used to process the original prediction sequence. Multi-scale exponential moving average smoothing is performed to obtain smoothing results at each scale. And determine the learnable weighting coefficients of the smoothing results. Next, learnable residual gating coefficients are introduced. For the original predicted sequence and smoothing results Perform dynamic weighted integration and output the final prediction result. .
[0026] Furthermore, the dynamic weighted integration takes the following form: .
[0027] According to a second aspect of the present invention, a coal mine load prediction system based on adaptive multi-scale EMA and multi-path signal feature fusion is provided, comprising the modules of the coal mine load prediction method based on adaptive multi-scale EMA and multi-path signal feature fusion as described above.
[0028] 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 method described in any one of the preceding descriptions.
[0029] According to a third 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.
[0030] The beneficial effects of this invention are:
[0031] To address the characteristics of coal mine load sequences, which exhibit strong volatility and coexistence of trend, periodicity, and short-term disturbances, this invention constructs a unified multidimensional input feature representation of complex working conditions through wavelet decomposition, K-Means clustering, trigonometric function encoding, and multi-path integration of two-level EMA. This multidimensional input feature compensates for the shortcomings of traditional methods in characterizing single-path models, providing trend information and sensitive disturbance cues for subsequent MF-AMES models. To address the problem of insufficient prediction stability during the alternating trend and disturbance phases, this invention employs a dual-layer LSTM (128 / 64 memory units) with functional division of labor for hierarchical modeling, and introduces an AMES module with learnable residual gating coefficients at the output end, achieving a dynamic adaptive balance between smoothing of the trend segment and agile response of the disturbance segment, thus ensuring both prediction accuracy and robustness under complex coal mine conditions. Attached Figure Description
[0032] Figure 1 This is a flowchart of the present invention.
[0033] Figure 2 This is a comparison chart of the load prediction values and actual values of the MF-AMES model proposed in this invention.
[0034] Figure 3 This is a comparison chart of the load forecasts and actual values from the DLinear model.
[0035] Figure 4 This is a comparison chart of the load forecasts and actual values from the Informer model.
[0036] Figure 5 This is a comparison chart of the load forecasts and actual values from the N-HiTS model. Detailed Implementation
[0037] 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.
[0038] Example 1: As Figures 1-5 As shown, according to a first aspect of the present invention, a coal mine load prediction method based on adaptive multi-scale EMA and multi-path signal feature fusion is provided, comprising:
[0039] S1. Obtain the raw time series data of coal mine electricity load and preprocess it to obtain the preprocessed time series data of electricity load.
[0040] S2. Sample window segmentation: The preprocessed electricity load time series data is segmented using a sliding window algorithm to obtain sample window data;
[0041] S3. Based on the sample window data, obtain wavelet decomposition features, clustering pattern features, time coding features, and smoothing features;
[0042] The S3 includes:
[0043] S3.1 Decompose the sample window data using Discrete Wavelet Transform (DWT) to obtain the wavelet decomposition features. ={ , , , The decomposition follows the following relationship:
[0044] ;
[0045] in, It is sample window data. It is the first Approximate components of the layer (reflecting low-frequency trends). It is the first Detailed components of the layer (reflecting high-frequency fluctuations);
[0046] In this invention, the following is set Based on the above formula, we obtain four components: , , , ,in Characterizing the long-term trends and process rhythms of coal mine production, while , , This characterizes short-term disturbance information such as equipment start-up and shutdown, sudden fluctuations, and high-frequency spikes at different scales;
[0047] S3.2, Wavelet decomposition features KMeans clustering was performed to obtain cluster labels that reflect the operating patterns at each time point. Clustering labels pass The encoding is converted into sparse vectors, and the sparse vectors are stacked and combined into a matrix in chronological order to form clustering pattern features. .
[0048] In the KMeans algorithm, the optimal number of clusters is determined using the elbow rule. That is, selecting the optimal number of clusters by minimizing the sum of squares within each cluster. The formula is as follows:
[0049] ;
[0050] in, Indicates the first Clusters, It is the center of the cluster. These are sample points within a cluster.
[0051] Cluster labels pass The process of converting the encoding into a sparse vector is as follows: Integer cluster labels at each time point are generated. Convert to a binary vector of length K, which is only valid at index equal to... The value in position 1 is 1, and the values in all other positions are 0.
[0052] S3.3, Wavelet decomposition features using sine and cosine functions The time and location information is processed to construct intraday and intraweek periodic time codes, forming intraday features. and intra-week characteristics By concatenating intraday and intraweek features, time-coded features are obtained. The formulas for intraday and weekly features are as follows:
[0053] ;
[0054] ;
[0055] ;
[0056] ;
[0057] in, Indicates a time step.
[0058] S3.4, Wavelet Decomposition Features Multi-scale exponential moving average (EMA) smoothing is performed to obtain smoothing characteristics. .
[0059] Specifically, two fixed smoothing factors (α=0.2 and α=0.4) are applied to the wavelet decomposition features for EMA calculation. The results of the EMA calculations for the two smoothing factors (α=0.2 and α=0.4) are then concatenated to obtain the smoothed features. The EMA calculation follows the following recursive relationship:
[0060] ;
[0061] in, Representative moment The values of wavelet components; As a smoothing factor, different For different smoothing scales, the smaller the value, the higher the weight of historical data and the smoother the sequence.
[0062] S4. Unify the wavelet decomposition features, clustering pattern features, time coding features, and smoothing features across the feature dimensions to generate multi-dimensional input features that can systematically characterize the complex working conditions of coal mines from different perspectives at multiple scales. ;in, ;
[0063] S5. Incorporate multidimensional input features After normalization, the input is fed into a double-layer stacked Long Short-Term Memory (LSTM) network for training. Specifically, S5 involves:
[0064] S5.1 Construct a double-layer stacked long short-term memory network; wherein the double-layer stacked long short-term memory network includes a first layer LSTM and a second layer LSTM; the first layer LSTM uses a large number of 128 memory units, focusing on capturing long-term dependencies to establish a stable long-term trend baseline; the second layer LSTM uses a smaller number of 64 memory units, focusing on short-term disturbance response to sense local rapid fluctuations.
[0065] S5.2, Inputting multidimensional features After normalization, the data is sequentially fed into a double-layer stacked Long Short-Term Memory network in time steps, and two corresponding hidden states are output. and The mathematical representation of this process is:
[0066] ;
[0067] ;
[0068] In the formula: and These are the first and second LSTM layers at time steps. The hidden state, and These are the hidden states from the previous moment.
[0069] S5.3, The hidden state sequence output by the second LSTM layer. The original predicted sequence is obtained by performing time-step mapping through a TimeDistributed fully connected layer. .
[0070] S6. Construct an adaptive multi-scale EMA fusion module (AMES); the adaptive multi-scale EMA fusion module is used to process the original prediction sequence. Multi-scale exponential moving average smoothing is performed (α=0.2 and α=0.4 in this embodiment of the invention) to obtain the smoothing results at each scale. And determine the learnable weighting coefficients of the smoothing results. Next, learnable residual gating coefficients are introduced. For the original predicted sequence and smoothing results Perform dynamic weighted integration and output the final prediction result. Its fusion form is as follows:
[0071] ;
[0072] This invention cleverly introduces residual gating coefficients. The adjustment mechanism can adaptively adjust the fusion ratio according to the degree of local fluctuation in the load sequence, so that the model can reduce the load during the trend stabilization phase. Weighting, focusing on smoothing prediction To maintain trend continuity; to improve during operating condition switching or short-term disturbance phases. Weights, focusing on the original predicted value This allows for dynamic adaptive balance.
[0073] According to a second aspect of the present invention, a coal mine load forecasting system based on adaptive multi-scale EMA and multipath signal feature fusion is provided, comprising modules of the coal mine load forecasting method based on adaptive multi-scale EMA and multipath signal feature fusion as described above. Each module in the above-described coal mine load forecasting system based on adaptive multi-scale EMA and multipath signal feature fusion can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.
[0074] 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 method described in any one of the above embodiments.
[0075] According to a third 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.
[0076] Example 2: This example provides a coal mine load forecasting method based on adaptive multi-scale EMA and multi-path signal feature fusion. The complete process is as follows: Figure 1 As shown, the specific steps include:
[0077] Step 1: Collect a batch of raw time-series data on coal mine electricity load from the Yunnan Province Real Coal Mine Monitoring System. The data covers 60 days, from September 19, 2024 to November 18, 2024, with a sampling interval of 15 minutes. Each day contains 96 load points, measured in kilowatts, totaling 5760 data points. The Yunnan Province Real Coal Mine Monitoring System monitors key equipment such as main ventilation fans and hoists, and their operating loads can effectively reflect the overall production conditions of the coal mine. The data contains various abnormal events (such as power surges and current anomalies), with abnormal days accounting for approximately 19%, thus exhibiting typical industrial time-series characteristics such as strong noise and category imbalance.
[0078] The raw time series data of coal mine electricity load were preprocessed as follows: First, missing values were filled with the mean and outliers were removed based on the 3σ principle; then, the uniformity of the time series was ensured by resampling at equal intervals, and all missing values were filled with time interpolation to ensure data continuity; finally, MinMaxScaler was used for normalization.
[0079] Step 2: The preprocessed electricity load time series data is segmented using a sliding window algorithm to generate sample window data for model training, validation, and testing. Each sample window is set to a length of 48 steps (WIN=48, corresponding to 12 hours of historical data input) to predict the load in the next step. Through the above steps, the preprocessed electricity load time series data is divided into three non-overlapping subsets in chronological order: a training set (70%), a validation set (15%), and a test set (15%).
[0080] Step 3: Perform a three-level discrete wavelet transform (DWT) on the sample window data, using the db4 wavelet basis. Decompose to obtain low-frequency approximate components. and three-layer high-frequency detail components , , There are four components in total, which constitute the wavelet decomposition features. .
[0081] Step 4: Apply the K-Means clustering algorithm to the wavelet decomposition features. Determine the optimal number of clusters using the elbow rule. (In this embodiment) The obtained pattern labels are converted into clustering pattern features using OneHot encoding. .
[0082] Step 5: Use sine and cosine functions to periodically encode the hours (period 24) and days of the week (period 7), forming a 4-dimensional time coding feature. .
[0083] Step 6: Employ the exponential moving average (EMA) method, introducing two fixed smoothing factors (e.g., ...). = 0.2, 0.4) for wavelet decomposition features Smoothing is performed to obtain smoothed features. .
[0084] Step 7, the above , , and The four types of features are uniformly concatenated along the feature dimension to generate multi-dimensional input features that can systematically represent complex working conditions. After normalization and window segmentation, the shape of the model input data is (number of samples, window length = 48, feature dimension).
[0085] Step 8, The input is fed into a two-layer stacked Long Short-Term Memory (LSTM) network for training. The first LSTM layer uses a large number of 128 memory units to focus on capturing long-term dependencies, while the second LSTM layer uses a smaller number of 64 memory units to focus on short-term perturbation responses.
[0086] Step 9: To enhance the model's generalization ability and prevent overfitting, a Dropout layer with a dropout rate of 0.2 was introduced after each LSTM layer. Finally, the original prediction sequence for each time step was obtained by mapping through a TimeDistributed fully connected layer. The model uses Adam as the optimizer and mean squared error (MSE) as the loss function.
[0087] Step 10: The model incorporates an Adaptive MultiScale EMA (AMES) fusion module at the output. This module is based on the original prediction sequence. Constructing EMA smoothing results with multiple different smoothing factors Simultaneously, the learnable weighting coefficients of the smoothing results are determined. This coefficient is applied to the trainable parameters through the Softmax function. get.
[0088] Step 11: The Adaptive MultiScale EMA (AMES) fusion module utilizes learnable residual gating coefficients. (From the trainable parameters inside the model) (obtained through mapping of the Sigmoid function) for the original prediction sequence. and smoothing results Perform dynamic weighted integration and output the final prediction result. .
[0089] Step 12: During model training, the EarlyStopping mechanism is used to monitor the validation set loss (val_loss), the patience value is set to 12 periods, and the optimal weights are restored to prevent model overfitting.
[0090] The implementation examples of this invention use three indicators—mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²)—to comprehensively evaluate the prediction performance.
[0091] To verify the effectiveness of this method, five representative load prediction models were selected as controls, including TFT, Informer, DLinear, N-HiTS and PatchTST.
[0092] Table 1 Comparison of prediction results of each model on the test set
[0093]
[0094] Table 1 shows that the MF-AMES model of this invention achieves optimal values across all indices, and its coefficient of determination is [missing information]. The R² value reached 0.9442, a significant improvement compared to the optimal baseline model PatchTST (R²=0.8860). The substantial reduction in RMSE (45.8365) indicates that this method exhibits higher robustness and lower error sensitivity when facing drastic fluctuations caused by working condition switching and short-term disturbances in coal mine loads. This invention fully demonstrates its advantages. Figures 2-5 The comparison of the predicted curves of the MF-AMES method of this invention with other comparative models over a full natural day of the test set with the actual load values (TFT, PatchTST, and Informer are all based on the transformer framework, so only Informer is shown). The comparison results of the predicted curves and the actual sequences in terms of phase and amplitude show that the prediction accuracy of this invention is higher than that of other models, which confirms that this method can provide reliable support for the accurate detection of abnormal electricity consumption behavior in coal mine safety production.
[0095] 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 coal mine load forecasting method based on adaptive multi-scale EMA and multi-path signal feature fusion, characterized in that, include: S1. Obtain the raw time series data of coal mine electricity load and preprocess it to obtain the preprocessed time series data of electricity load. S2. Sample window segmentation: The preprocessed electricity load time series data is segmented using a sliding window algorithm to obtain sample window data; S3. Obtain wavelet decomposition features based on the sample window data. Clustering pattern characteristics Time coding features and smooth features ; S4. Consolidate the wavelet decomposition features, clustering pattern features, time-coding features, and smoothing features along the feature dimension to generate multi-dimensional input features. ;in, ; S5. Construct a double-layer stacked long short-term memory network; incorporate multi-dimensional input features. After normalization, it is used as the input of a stacked long short-term memory network. The hidden state output by the second LSTM is mapped step-by-step through a fully connected layer to obtain the original prediction sequence. S6. Construct an adaptive multi-scale EMA fusion module; use the original prediction sequence as input to the adaptive multi-scale EMA fusion module and output the final prediction result.
2. The coal mine load forecasting method based on adaptive multi-scale EMA and multi-path signal feature fusion as described in claim 1, characterized in that, The S3 includes: S3.1 Decompose the sample window data using discrete wavelet transform to obtain wavelet decomposition features. ={ , , , The decomposition follows the following relationship: ； in, It is sample window data. It is the first Approximate components of the layer, It is the first Layer detail components; set up Based on the above formula, we obtain four components: , , , ,in Characterizing the long-term trends and process rhythms of coal mine production, while , , This characterizes equipment start-up and shutdown, sudden fluctuations, and short-term disturbances with high-frequency spikes at different scales; S3.2, Wavelet decomposition features KMeans clustering was performed to obtain cluster labels that reflect the operating patterns at each time point. Clustering labels pass The encoding is converted into sparse vectors, and the sparse vectors are stacked and combined into a matrix in chronological order to form clustering pattern features. ; S3.3, Wavelet decomposition features using sine and cosine functions The time and location information is processed to construct intraday and intraweek periodic time codes, forming intraday features. and intra-week characteristics By concatenating intraday and intraweek features, time-coded features are obtained. ; S3.4, Wavelet Decomposition Features Multi-scale exponential moving average smoothing is performed to obtain smoothing features. .
3. The coal mine load forecasting method based on adaptive multi-scale EMA and multi-path signal feature fusion according to claim 2, characterized in that, In the KMeans algorithm, the optimal number of clusters is determined using the elbow rule. ; Cluster labels pass The process of converting the encoding into a sparse vector is as follows: Integer cluster labels at each time point are generated. Convert to a binary vector of length K, which is only valid at index equal to... The value in position 1 is 1, and the values in all other positions are 0.
4. The coal mine load forecasting method based on adaptive multi-scale EMA and multi-path signal feature fusion according to claim 1, characterized in that, In S5, a Dropout layer with a dropout rate of 0.2 is introduced after each LSTM layer.
5. The coal mine load forecasting method based on adaptive multi-scale EMA and multi-path signal feature fusion according to claim 1, characterized in that, The adaptive multi-scale EMA fusion module is used to process the original prediction sequence. Multi-scale exponential moving average smoothing is performed to obtain smoothing results at each scale. And determine the learnable weighting coefficients of the smoothing results. Next, learnable residual gating coefficients are introduced. For the original predicted sequence and smoothing results Perform dynamic weighted integration and output the final prediction result. .
6. The coal mine load forecasting method based on adaptive multi-scale EMA and multi-path signal feature fusion according to claim 5, characterized in that, The dynamic weighted integration form is as follows: .
7. A coal mine load forecasting system based on adaptive multi-scale EMA and multi-path signal feature fusion, characterized in that, A module comprising the method of any one of claims 1-6.
8. A processor, characterized in that, The processor is used to run a program, wherein the program, when running, performs the steps of the method according to any one of claims 1-6.
9. 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 1-6.