Electric coal inventory prediction method and device, electronic equipment and storage medium
By using a multi-source time-series data acquisition and adaptive learning method, combined with a multi-scale time-series prediction model and a dual neural network, the problems of feature extraction and dynamic adaptation in thermal coal inventory prediction were solved, achieving more efficient thermal coal inventory prediction.
Patent Information
- Application Number
- CN202512005452.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-29
- Publication Date
- 2026-05-15
AI Technical Summary
Existing methods for predicting thermal coal inventory are insufficient in terms of effective feature extraction and dynamic adaptability, making it difficult to handle complex multi-scale features and data distribution changes, resulting in low prediction accuracy.
A multi-source time series data acquisition method is adopted. By adaptively learning normalization and denormalization parameters, a multi-scale time series prediction model is used to extract and fuse features at different time scales. A closed-loop prediction system is formed by combining dual neural networks and multiple predictors.
It improves the accuracy and reliability of thermal coal inventory forecasting, better copes with data fluctuations and distribution drift, and enhances the model's ability to capture complex time-series changes.
Smart Images
Figure CN122047584A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of inventory management technology, and in particular to a method, apparatus, electronic device, and storage medium for predicting thermal coal inventory. Background Technology
[0002] Inventory management at coal-fired power plants is crucial for ensuring power supply and economic operation. my country's coal resources are mainly distributed in the north and west, while power load centers are concentrated in the eastern coastal areas. This imbalance between production and demand leads to coal reliance on long-distance transportation, making it susceptible to factors such as transportation capacity and weather, resulting in high supply chain risks. Power plant inventory management needs to strike a balance between supply security and cost control. Inventory-related costs account for more than 70% of total power generation costs; therefore, improving the accuracy of inventory forecasting is of practical urgency.
[0003] Existing prediction methods can be mainly divided into the following categories: traditional time series methods such as ARIMA and exponential smoothing, which use historical data to fit trends and seasonality, but are not adaptable to complex multi-scale features and changes in data distribution; machine learning methods such as support vector machines and random forests, which can handle nonlinear relationships, but rely on manual feature selection and have limited robustness to multi-dimensional features and data distribution drift; and deep learning methods such as LSTM and GRU, which can automatically extract time-dependent features, but still have shortcomings in capturing multi-scale information and dealing with data distribution drift.
[0004] Overall, existing methods still have room for improvement in terms of effective feature extraction and dynamic adaptability. Summary of the Invention
[0005] The purpose of this invention is to provide at least one method, apparatus, electronic device, and storage medium for predicting thermal coal inventory, which can at least solve the technical problems of insufficient feature extraction and dynamic adaptation capabilities of existing methods, and at least achieve more efficient extraction of time-dependent features, thereby improving the accuracy of thermal coal inventory prediction.
[0006] To address the aforementioned technical problems, at least one embodiment of this application provides a method for predicting thermal coal inventory, comprising: Acquire multi-source time-series data related to thermal coal inventory; Based on the fluctuation patterns of the multi-source time series data, adaptive learning is used to obtain the normalized parameters of the historical window and the inverse normalized parameters of the prediction window. The normalization parameters are used to normalize the multi-source time series data, and the normalized data is input into a preset multi-scale time series prediction model to obtain preliminary prediction results. The multi-scale time series prediction model is used to extract and fuse time series features of different time scales from the normalized data, and to make predictions based on the fused multi-scale time series features to obtain the preliminary prediction results. The preliminary prediction results are inversely normalized using the inverse normalization parameter to obtain the final predicted value of thermal coal inventory.
[0007] At least one embodiment of this application also provides a device for predicting coal inventory. The acquisition module is used to acquire multi-source time-series data related to thermal coal inventory; The normalization parameter generation module is used to adaptively learn the normalization parameters of the historical window and the inverse normalization parameters of the prediction window based on the fluctuation patterns of the multi-source time series data. The prediction module is used to perform normalization transformation on the multi-source time series data using the normalization parameters, and input the normalized data into a preset multi-scale time series prediction model to obtain preliminary prediction results. The multi-scale time series prediction model is used to extract and fuse time series features of different time scales from the normalized data, and make predictions based on the fused multi-scale time series features to obtain the preliminary prediction results. The inverse normalization module is used to inverse normalize the preliminary prediction results using the inverse normalization parameters to obtain the final predicted value of thermal coal inventory.
[0008] At least one embodiment of this application also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described method for predicting coal inventory.
[0009] At least one embodiment of this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting coal inventory.
[0010] The methods, apparatus, electronic devices, and storage media for predicting thermal coal inventory provided in this application comprehensively cover various factors affecting thermal coal inventory by acquiring multi-source time-series data, laying the foundation for accurate prediction. Adaptive learning of the normalized parameters of the historical window and the denormalized parameters of the prediction window effectively addresses data fluctuations and improves data processing adaptability. Utilizing a multi-scale time-series prediction model to extract and fuse features from different time scales fully uncovers the inherent patterns in the data, enhancing the model's ability to capture complex time-series changes, thereby obtaining more accurate preliminary prediction results. Finally, denormalization yields the final predicted value. The entire process is logically rigorous and effectively improves the accuracy and reliability of thermal coal inventory prediction.
[0011] In some optional embodiments, the adaptive learning of the normalized parameters of the historical window and the denormalized parameters of the prediction window based on the fluctuation patterns of the multi-source time series data includes: By using dual neural networks, based on the same historical time series data, the normalized parameters of the historical window and the inverse normalized parameters of the prediction window are learned separately.
[0012] In this embodiment, dual neural networks learn normalized and denormalized parameters separately, allowing each neural network to focus on a single task. This avoids interference and confusion that can occur when a single network processes two related but different tasks simultaneously. This clear division of labor helps each network better optimize its parameters, improves learning efficiency, and thus enhances the overall training performance of the model.
[0013] In some optional embodiments, the multi-source time-series data is normalized according to the following formula:
[0014] in, It is normalized data. It is historical time-series data. and These are affine parameters. and yes Time series The distribution of parameters in the input space describes the parameters.
[0015] In this embodiment, affine parameters are used. and The parameters are described by combining the distribution of time series i at time t in the input space. and It can adaptively normalize the data based on the actual fluctuations. This makes the normalization process more closely match the true characteristics of the data and effectively handle fluctuations of different scales and patterns.
[0016] In some optional embodiments, the preliminary prediction results are denormalized according to the following formula:
[0017] in, The multi-scale time series prediction model is for the future The predicted value for each time step, i.e., the preliminary prediction result, This is the final predicted value of thermal coal inventory after denormalization. and yes Time series The parameters are distributed in the output space.
[0018] In this embodiment, the distribution description parameters of time series i at time t in the output space are considered through precise inverse normalization calculation. and and affine parameters and This can minimize errors caused by data conversion, ensure that the obtained coal inventory forecast is as close as possible to the actual value, and improve the reliability of the forecast.
[0019] In some optional embodiments, temporal features at different time scales are extracted and fused from the normalized data, including: The normalized data is decomposed into components representing periodic fluctuations and components representing long-term trends. For the periodic fluctuation component, a first-direction fusion operation is performed to fuse high-frequency scale information into low-frequency scale information; For the long-term trend component, a second-direction fusion operation is performed to fuse low-frequency scale information into high-frequency scale information. The periodic fluctuation components and the long-term trend components are added together after fusion to obtain the fused multi-scale time series features.
[0020] In this embodiment, the data is decomposed into components representing periodic fluctuations and components representing long-term trends. This allows for in-depth analysis of these different types of fluctuations, avoiding interference between different features, and thus capturing various information in the data more comprehensively and meticulously. By performing fusion operations in different directions on the periodic fluctuation components and the long-term trend components, it is possible to better adapt to changes in the data at different time scales.
[0021] In some optional embodiments, prediction is performed based on the fused multi-scale temporal features to obtain the preliminary prediction result, including: Multiple parallel sub-predictors are used to predict the feature correspondences at different scales in the fused multi-scale temporal features respectively; The outputs of each of the sub-predictors are integrated to obtain the preliminary prediction result.
[0022] In this embodiment, multiple parallel sub-predictors can be specifically designed and optimized for features at different scales, giving full play to the unique role of each scale feature in prediction, thereby capturing inventory change patterns more accurately from multiple perspectives and improving overall prediction accuracy.
[0023] In some optional embodiments, the loss function of the multi-scale time series prediction model includes a main loss term based on the error between the predicted and true values, and a regularization term used to constrain the learning effect of the inverse normalization parameters.
[0024] In this embodiment, the regularization term constrains the learning effect of the inverse normalization parameters, ensuring that the inverse normalization process accurately restores the model output to the original data scale. This helps avoid significant deviations between the prediction results and the actual situation due to unreasonable inverse normalization parameters, further improving the accuracy and reliability of the final prediction value. Attached Figure Description
[0025] One or more embodiments are illustrated by way of example with reference to the accompanying drawings, and these illustrative descriptions do not constitute a limitation on the embodiments.
[0026] Figure 1 This is a flowchart of a method for predicting thermal coal inventory provided in one embodiment of this application; Figure 2 This is a schematic diagram of the architecture of the MTDNet model provided in one embodiment of this application; Figure 3 This is a schematic diagram of a coal inventory prediction device provided in another embodiment of this application. Detailed Implementation
[0027] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the various embodiments of this application to help readers better understand this application. However, the technical solutions claimed in this application can be implemented even without these technical details and various changes and modifications based on the following embodiments. The division of the various embodiments below is for the convenience of description and should not constitute any limitation on the specific implementation of this application. The various embodiments can be combined with and referenced by each other without contradiction.
[0028] To facilitate understanding of the embodiments of this application, relevant content regarding the forecasting of existing thermal coal inventory will be introduced first.
[0029] China possesses abundant coal resources, and coal for power generation is a crucial component of its energy system. Due to China's energy structure—abundant coal resources but low self-sufficiency in oil and natural gas—coal is unlikely to be replaced by emerging energy sources in the short term and will remain the primary source of thermal power generation. However, China's coal resources are mainly concentrated in the northern and western regions, where coal production is highly disproportionate to local economic prosperity and demand. This imbalance between coal resource distribution and regional demand necessitates cross-regional allocation from the main coal-producing areas in the north and west to meet the demand of power load centers along the eastern coast. This process is constrained by complex factors such as transportation capacity and extreme weather; supply chain disruptions could trigger regional power crises. For coal-fired power plants, inventory management requires a balance between supply security and economic efficiency—insufficient inventory threatens the execution of power generation plans, while excessive inventory leads to calorific value loss and wasted transportation capacity. Statistics show that over 70% of power plant generation costs are related to coal inventory, highlighting the urgent need to improve forecasting accuracy to support optimized decision-making.
[0030] Existing technologies similar to this invention in the field of coal inventory forecasting can be categorized into those based on traditional time series analysis, machine learning algorithms, deep learning algorithms, and forecasting schemes combining prior knowledge and simple preprocessing. Traditional methods such as ARIMA (Autoregressive Moving Average) and exponential smoothing primarily utilize the time series characteristics of historical inventory data to build models for forecasting. They identify trends and seasonal patterns in the data by fitting it to past data, thereby estimating future inventory levels. However, these methods often exhibit limitations when dealing with complex multi-scale characteristics and situations where data distribution changes, making it difficult to adapt to the dynamic characteristics of coal inventory data.
[0031] Machine learning algorithms such as Support Vector Machines (SVM) and Random Forests can handle non-linear relationships by being trained on large amounts of historical data to learn the mapping relationship between features in the data and inventory. However, in terms of feature selection, manual feature selection or reliance on specific feature selection algorithms is usually required. For data with massive amounts of multi-dimensional features, it is difficult to effectively select the features most valuable for prediction, and the adaptability and robustness of the model need to be improved when faced with data distribution drift.
[0032] Deep learning methods such as LSTM (Long Short-Term Memory) or GRU (Gated Recurrent Unit) can automatically extract long-term and short-term features from time-series data, and have certain advantages in processing time-dependent coal inventory data. However, LSTM and GRU are relatively weak in multi-scale feature extraction and are quite sensitive to changes in data distribution. When the data distribution drifts, the predictive performance of the model is easily affected.
[0033] Existing methods suffer from limitations when handling features across different time scales. For instance, thermal coal inventory data exhibits short-term intraday fluctuations, long-term seasonal trends, and year-end cyclical patterns. Traditional forecasting models perform poorly in simultaneously extracting and fusing these multi-scale features, struggling to effectively couple features across different scales. This results in an incomplete and superficial understanding of the data, impacting forecast accuracy. Furthermore, statistical properties such as mean and variance in time series data frequently change over time, leading to data distribution drift. This temporal distribution variation is one of the major challenges hindering accurate time series forecasting.
[0034] To address the technical problems of insufficient effective feature extraction and dynamic adaptability in existing methods, this invention proposes a method for predicting coal inventory. The implementation details of the coal inventory prediction method in this embodiment are described below. The following content is only for ease of understanding and is not necessary for implementing this solution.
[0035] Example 1: The coal inventory forecasting method of this embodiment can be applied to electronic devices with communication, computing, and data storage capabilities. Its specific process can be as follows: Figure 1 As shown, it includes: Step 110: Obtain multi-source time-series data related to thermal coal inventory; Specifically, thermal coal inventory refers to the inventory of coal used for power generation. Thermal coal inventory is affected by various factors, therefore, it is necessary to collect relevant time-series data from multiple different data sources. Multi-source time-series data can include production records from coal enterprises, electricity demand data from power companies, logistics data from the transportation sector, market data, etc. The collected data should have time-series characteristics, with data points recorded in chronological order to facilitate analysis of trends and patterns over time.
[0036] Step 120: Based on the fluctuation pattern of the multi-source time series data, adaptive learning is used to obtain the normalized parameters of the historical window and the inverse normalized parameters of the prediction window. Specifically, fluctuation patterns refer to the patterns and trends of change exhibited by data over time, including periodic fluctuations, trend changes, and random fluctuations. Because the fluctuation patterns of multi-source time series data are complex and diverse, and may be influenced by various factors such as seasons, market conditions, and unexpected events, adaptive learning methods are needed to determine appropriate normalization and denormalization parameters.
[0037] Normalization parameters are used to normalize raw data. Through specific mathematical transformations, the data is scaled to a specified range. Common normalization methods include min-max normalization and Z-score normalization. Inverse normalization parameters, corresponding to normalization parameters, are used to restore the normalized data to its original scale. This is achieved through the inverse mathematical transformation. Adaptive learning is a machine learning method that automatically adjusts its parameters and structure based on the characteristics and changes of the input data to better adapt to the data, improving model performance and generalization ability.
[0038] For example, adaptive learning methods can dynamically analyze recent historical data (such as the past 30 days) used for training and input, and independently calculate scaling parameters such as the mean and standard deviation within that window. Considering that future prediction windows (such as the following week) may have different data distributions than historical windows, adaptive learning methods can be used to estimate or learn the corresponding parameters for future windows, which can then be used to restore the prediction results to the true scale.
[0039] This step aims to address the modeling challenges posed by data non-stationarity (such as trends and seasonality). Traditional methods typically use fixed normalization parameters (such as global maximum / minimum values) for the entire dataset, which performs poorly when the data distribution changes over time. The core of this embodiment lies in "adaptive learning," which automatically and dynamically determines the processing parameters (in this case, normalization parameters) based on the statistical characteristics of the input data itself, rather than using pre-set fixed values, thereby enhancing the model's ability to adapt to dynamic changes in the data.
[0040] Step 130: The normalization parameters are used to perform normalization transformation on the multi-source time series data, and the normalized data is input into a preset multi-scale time series prediction model to obtain preliminary prediction results. The multi-scale time series prediction model is used to extract and fuse time series features of different time scales from the normalized data, and to make predictions based on the fused multi-scale time series features to obtain the preliminary prediction results. Specifically, this step aims to overcome the limitations of traditional single-model approaches in capturing complex time patterns. After data normalization, the processed data is input into a multi-scale time series forecasting model. The core task of this model is to extract valuable information from the normalized data. First, it extracts time series features at different time scales, such as short-term (daily, weekly) fluctuations and long-term (monthly, quarterly, yearly) trend features. Then, a specific fusion method is used to fuse these features at different scales, enabling the model to comprehensively consider data changes across various time scales. Finally, based on the fused multi-scale time series features, a preliminary forecast of thermal coal inventory is output.
[0041] Step 140: Use the inverse normalization parameter to inverse normalize the preliminary prediction result to obtain the final predicted value of thermal coal inventory.
[0042] Specifically, since the model's initial predictions are based on normalized data, their numerical range may differ from the original coal inventory data. Therefore, it is necessary to use the previously learned inverse normalization parameters to inverse normalize the initial predictions, restoring the predicted values to the scale of the original data.
[0043] The method flow of this embodiment constitutes a complete, closed-loop prediction system from data integration -> adaptive preprocessing -> multi-scale modeling -> result postprocessing. It addresses data drift through adaptive parameter learning and utilizes multi-scale models to deeply mine complex time series patterns, aiming to improve the accuracy of thermal coal inventory prediction in dynamic and complex environments.
[0044] In summary, this embodiment, by acquiring multi-source time-series data, comprehensively covers various factors affecting thermal coal inventory, laying the foundation for accurate prediction. Adaptive learning of the normalized parameters of the historical window and the denormalized parameters of the prediction window effectively addresses data fluctuations and improves data processing adaptability. Utilizing a multi-scale time-series prediction model to extract and fuse features from different time scales fully uncovers the inherent patterns in the data, enhancing the model's ability to capture complex time-series changes, thereby obtaining more accurate preliminary prediction results. Finally, denormalization yields the final predicted value. The entire process is logically rigorous and effectively improves the accuracy and reliability of thermal coal inventory prediction.
[0045] In some optional embodiments, the adaptive learning of the normalized parameters of the historical window and the denormalized parameters of the prediction window based on the fluctuation patterns of the multi-source time series data includes: using a dual neural network to learn the normalized parameters of the historical window and the denormalized parameters of the prediction window based on the same historical time series data.
[0046] Specifically, a Double-Net framework is introduced to adaptively capture the distribution parameters of data in the historical window and the prediction window from the input sequence, thereby forming a preprocessing method that can effectively capture the distribution changes of data.
[0047] Historical data, input into Back-Net and Forward-Net respectively, can be represented as follows: (1) (2) in, and yes Time series The distribution of parameters in the input space is used in the normalization process; and yes Time series The distribution of parameters in the input and output spaces is used in the inverse normalization process, where N is the number of time series. This is the historical time-series data input to Back-Net and Forward-Net.
[0048] Although Back-Net and Forward-Net use the same input data, they have different learning objectives. Furthermore, Back-Net and Forward-Net can be selected based on the prediction task, offering flexibility and robustness. In this prediction task, Back-Net and Forward-Net are fully connected neural networks.
[0049] In this embodiment, dual neural networks learn normalized and denormalized parameters separately, allowing each neural network to focus on a single task. This avoids interference and confusion that can occur when a single network processes two related but different tasks simultaneously. This clear division of labor helps each network better optimize its parameters, improves learning efficiency, and thus enhances the overall training performance of the model.
[0050] In some optional embodiments, the multi-source time-series data is normalized according to the following formula: (3) in, It is normalized data. and These are affine parameters. and yes Time series The distribution of parameters in the input space describes the parameters.
[0051] Specifically, the algorithm in this embodiment combines affine learning and normalization, and uses the distribution of the input space to describe the parameter normalization time series. This can stabilize the time series while taking into account the non-stationary information in the series, so that the normalized data contains more original sequence features.
[0052] In some optional embodiments, extracting and fusing time-series features at different time scales from the normalized data includes: decomposing the normalized data into components representing periodic fluctuations and components representing long-term trends; for the periodic fluctuation components, performing a first-direction fusion operation that fuses high-frequency scale information to low-frequency scales; for the long-term trend components, performing a second-direction fusion operation that fuses low-frequency scale information to high-frequency scales; and adding the fused periodic fluctuation components and long-term trend components to obtain the fused multi-scale time-series features.
[0053] Specifically, during the decomposition process, due to the coupling of multiple features, thermal coal inventory may exhibit superimposed periodic characteristics of short-term, medium-term, and long-term periods. Therefore, Fourier transform is used to capture and separate these superimposed periodic components. The main steps include: converting the sequence from the time domain to the frequency domain, calculating the amplitude of each frequency component, setting the DC component amplitude to zero to eliminate the influence of the trend term, and then taking the most significant periodic component in the signal to approximate the seasonal term. In this embodiment, the first three frequency components with the largest amplitudes are selected, and then the selected frequency components are reconstructed into a seasonal signal in the time domain through inverse Fourier transform. Finally, the trend term is obtained by subtracting the seasonal term from the original sequence.
[0054] In some alternative embodiments, more advanced feature selection algorithms can be used for feature selection, such as combining intelligent optimization algorithms based on genetic algorithms, particle swarm optimization algorithms, and other intelligent optimization algorithms with feature selection methods to improve the accuracy and effectiveness of feature selection.
[0055] In the process of mixing seasonal and trend terms, the data can be described using periodic features at different scales. Detailed features at smaller scales in the seasonal term can provide richer information for larger-scale seasonal modeling. Therefore, when fusing the seasonal term, high-frequency time information is merged upwards into low-frequency information (i.e., the first-direction fusion operation). Trend term modeling focuses on the clarity of macroscopic information; detailed changes may introduce noise when capturing macroscopic trends. Since low-frequency time series more easily provide clear macroscopic information, when fusing the trend term, low-frequency trend information is merged downwards into high-frequency information (i.e., the second-direction fusion operation). Finally, the fused trend term is added to the seasonal term to obtain a multi-scale sequence after the interaction of information from different time scales, i.e., the fused multi-scale time series features.
[0056] In some optional embodiments, when performing multi-scale feature fusion, a network structure with multiple branches can be constructed, with each branch responsible for processing features at a specific scale. The outputs of each branch are then fused at the end. Alternatively, a dedicated feature fusion network can be designed, taking features at different scales as input and gradually fusing and abstracting features through a series of operations such as convolution, pooling, and activation. The final output is the fused feature used for prediction. In coal inventory prediction, a multi-layered feature fusion network can be constructed based on task requirements and data characteristics to effectively fuse and model multi-scale features, thereby improving prediction accuracy.
[0057] In this embodiment, the data is decomposed into components representing periodic fluctuations and components representing long-term trends. This allows for in-depth analysis of these different types of fluctuations, avoiding interference between different features, and thus capturing various information in the data more comprehensively and meticulously. By performing fusion operations in different directions on the periodic fluctuation components and the long-term trend components, it is possible to better adapt to changes in the data at different time scales.
[0058] In some optional embodiments, the preliminary prediction result is obtained by making predictions based on the fused multi-scale temporal features, including: using multiple parallel sub-predictors to predict the feature correspondences of different scales in the fused multi-scale temporal features respectively; and integrating the output results of each sub-predictor to obtain the preliminary prediction result.
[0059] Specifically, multiple independent predictors (such as small neural networks) are used in parallel to predict multi-scale temporal features. Each predictor can focus on learning patterns at a specific time scale, thus learning more accurately. Integrating the results of multiple predictors (e.g., by averaging or weighted averaging) can smooth out the errors of individual predictors, reduce the risk of overfitting, and improve the stability of the final prediction.
[0060] In this embodiment, multiple parallel sub-predictors can be specifically designed and optimized for features at different scales, giving full play to the unique role of each scale feature in prediction, thereby capturing inventory change patterns more accurately from multiple perspectives and improving overall prediction accuracy.
[0061] In some optional embodiments, the loss function of the multi-scale time series prediction model includes a main loss term based on the error between the predicted and true values, and a regularization term used to constrain the learning effect of the inverse normalization parameters.
[0062] Specifically, due to the distribution offset between the input and output spaces, Forward-Net may find it more difficult to learn the data distribution parameters in the output space, and this difficulty becomes more pronounced as the prediction window increases. To address this, some prior knowledge can be introduced into the loss function to enhance the learning performance of Forward-Net during training.
[0063] Assume that the multi-scale time series prediction model uses the classic mean absolute error (MAE) as the model loss function, as shown in Equation (4).
[0064] (4) in, It is the batch size. It is the number of time series. It is any point in time where samples are randomly collected from a batch of samples. It is a model pair back The predicted value at each time step, To predict the true value corresponding to the data. Although the statistical characteristics of the mean cannot fully reflect the distribution, it still has certain output space characteristics. Therefore, the loss function can be optimized using prior knowledge as follows: (5) in, It represents the weight of the prior part; The data size parameter for the output of Forward-Net. The closer the Forward-Net is to the mean of the prediction space, the better its learning performance is to some extent.
[0065] In this embodiment, the regularization term constrains the learning effect of the inverse normalization parameters, ensuring that the inverse normalization process accurately restores the model output to the original data scale. This helps avoid significant deviations between the prediction results and the actual situation due to unreasonable inverse normalization parameters, further improving the accuracy and reliability of the final prediction value.
[0066] In summary, the multi-scale time series prediction model in this embodiment has the following characteristics: At the feature level: through Fourier decomposition and directional multi-scale fusion, information-rich feature representations are intelligently constructed; At the prediction level: Multiple predictor integration is used to robustly generate the final results; At the training level: By designing a loss function based on prior knowledge, not only is the prediction accuracy optimized, but the key step of data preprocessing (Forward-Net) is also actively optimized, enabling the entire system to form a closed loop and evolve together.
[0067] This combination of technologies makes multi-scale time series forecasting models particularly suitable for complex industrial time series forecasting tasks such as thermal coal inventory, which are highly periodic, multi-scale, and whose distribution may vary over time.
[0068] In some optional embodiments, the preliminary prediction results are denormalized according to the following formula: (6) in, The multi-scale time series prediction model is for the future The predicted value for each time step, i.e., the preliminary prediction result, This is the final predicted value of thermal coal inventory after denormalization. and yes Time series The parameters are distributed in the output space.
[0069] In this embodiment, the distribution description parameters of time series i at time t in the output space are considered through precise inverse normalization calculation. and and affine parameters and This can minimize errors caused by data conversion, ensure that the obtained coal inventory forecast is as close as possible to the actual value, and improve the reliability of the forecast.
[0070] In some alternative embodiments, the superiority of the method proposed in this embodiment in predicting thermal coal inventory can be determined using the following calculation formula: (7) in, Indicates the sequence length. and These are the true value and the predicted value of the test set, respectively, used to compare the ratio of the model's prediction error to the true value. Their values range from 0 to positive infinity, with smaller values indicating better prediction performance.
[0071] Example 2: Based on the above embodiments, this embodiment provides an application example of a method for predicting thermal coal inventory. Addressing the problems of weak multi-scale coupling processing capability and significant susceptibility to data distribution drift in existing technologies, the method employs an optimized RevIn algorithm to adaptively capture the distribution changes of data between historical and prediction windows based on dynamic data change patterns, and scales the model prediction results to the distribution state corresponding to the future window to reduce the impact of data distribution drift; the prediction uses the MTDNet model.
[0072] The MTDNet model separates the trend and seasonal components of the downsampled series at each scale using a past information decomposition and mixing module. This isolates the impact of short-term fluctuations on long-term trends, allowing for a more refined capture of long-term fluctuation trends in the time series. Then, it separately mixes the seasonal and trend components at different scales to allow for information interaction across multiple scales, enhancing the model's ability to capture features at different time scales. A multi-predictor mixing module uses multiple predictors to predict subsequences at each scale, and integrates the results to generate the final predicted value. This fully utilizes the complementarity of information at different scales, ultimately achieving good accuracy in predicting thermal coal inventory.
[0073] The specific implementation steps include the following: S1: Collect and organize data related to the supply chain network of thermal coal inventory, temperature, and power plant generation. Fill in missing values using interpolation and remove outliers using the three-standard-deviation method to obtain cleaned data.
[0074] Specifically, by collecting historical data on coal reserves, coal intake, coal consumption, power generation, temperature, freight rates, and coal prices from power plants, and then preprocessing, cleaning, and filling the data, the statistical resolution of each feature is unified to the day.
[0075] S2: Divide the dataset into training and testing sets. The training and testing sets should be divided in an 8:2 ratio.
[0076] S3: Downsampled training data, normalized using the RevIn algorithm, and then fed into the MTDNet model.
[0077] Specifically, after downsampling, the optimized RevIn algorithm learns normalized parameters based on the dynamic changes in the data. This algorithm combines normalization with affine learning and introduces a Double-Net framework to adaptively capture the distribution parameters of the data in the historical and prediction windows from the input sequence, thus forming a preprocessing method that can effectively capture changes in data distribution. 1) Learning normalization parameters: Historical data is input into Back-Net and Forward-Net respectively, which can be represented as: (1) (2) in, and yes Time series The distribution of parameters in the input space is used in the normalization process; and yes Time series The distribution of parameters in the input and output spaces is used in the inverse normalization process, where N is the number of time series. This is the historical time-series data input to Back-Net and Forward-Net.
[0078] Although Back-Net and Forward-Net use the same input data, they have different learning objectives. Furthermore, Back-Net and Forward-Net can be selected based on the prediction task, offering flexibility and robustness. In this prediction task, Back-Net and Forward-Net are fully connected neural networks.
[0079] 2) Normalization: The algorithm combines affine learning and normalization, and uses the distribution of the input space to describe the parameter normalization time series, as shown in formula (3).
[0080] (3) in, It is normalized data. and These are two learnable affine parameters. and The output is obtained from Back-Net. This process can stabilize the time series while taking into account the non-stationary information in the series, so that the normalized data contains more of the original sequence features.
[0081] S4: The MTDNet model uses a past information decomposition and fusion module to decompose and fuse seasonal and trend terms. Then, the fused features at each scale are fed into a multi-prediction value fusion module for multi-scale prediction and synthesis of the final prediction results. Prior knowledge terms related to the RevIn algorithm are added to the loss function to improve the performance of the RevIn algorithm, and the loss function is used to optimize the neural network.
[0082] Specifically, such as Figure 2The diagram shows the specific architecture of the MTDNet model. The MTDNet model includes a multi-scale downsampling module, a past information decomposition and mixing module, and a future multi-prediction value mixing module. The output of the multi-scale downsampling module is normalized using the RevIn algorithm and then input into the past information decomposition and mixing module. During the decomposition process, due to multi-feature coupling, the thermal coal inventory may exhibit multiple periodic superposition characteristics, including short-term, medium-term, and long-term periods. Therefore, Fourier transform is used to capture and separate these superimposed periodic components. The main steps are: converting the sequence from the time domain to the frequency domain, calculating the amplitude of each frequency component, setting the DC component amplitude to zero to eliminate the influence of the trend term, then taking the most significant periodic component in the signal to approximate the seasonal term (here, the three frequency components with the largest amplitudes are selected), and then reconstructing the seasonal signal in the time domain using inverse Fourier transform. Finally, the trend term is obtained by subtracting the seasonal term from the original sequence.
[0083] In the process of mixing seasonal and trend terms, research shows that it can be described by periodic features at different scales. Detailed features at smaller scales in the seasonal term can provide richer information for larger-scale seasonal modeling. Therefore, when fusing the seasonal term, high-frequency time information is merged upwards into low-frequency information. Trend term modeling focuses on the clarity of macroscopic information; detailed changes may introduce noise when capturing macroscopic trends. Since low-frequency time series are more likely to provide clear macroscopic information, low-frequency trend information is merged downwards into high-frequency information when fusing the trend term. Finally, the fused trend term is added to the seasonal term to obtain a multi-scale sequence after the interaction of information from different time scales, which serves as the output of the past information decomposition and mixing module. The mixing structure is as follows: Figure 1 As shown in the bottom right corner The multi-predictor hybrid module uses multiple predictors to predict subsequences at each scale, and integrates the results of these predictors to generate the final predicted value.
[0084] Because of the distribution offset between the input and output spaces, Forward-Net may find it more difficult to learn the distribution parameters of the output space data, and this difficulty becomes more pronounced as the prediction window increases. To address this, some prior knowledge can be introduced into the loss function to enhance the learning performance of Forward-Net during training.
[0085] Assume the model uses the classic Mean Absolute Error (MAE) as the model loss function, as shown in Equation (4): (4) in, It is the batch size. It is the number of time series. It is any point in time where samples are randomly collected from a batch of samples. It is a model pair back The predicted value at each time step, To predict the true value corresponding to the data. Although the statistical characteristics of the mean cannot fully reflect the distribution, it still has certain output space characteristics. Therefore, the loss function can be rewritten using prior knowledge as follows: (5) in, It represents the weight of the prior part; The data size parameter for the output of Forward-Net. The closer the Forward-Net is to the mean of the prediction space, the better its learning performance is to some extent.
[0086] S5: The output of the MTDNet model is denormalized using the improved RevIn algorithm to obtain the final prediction result.
[0087] Specifically, the output of the MTDNet model is denormalized using the improved RevIn algorithm to obtain the final prediction result, as shown in Equation (6): (6) in, It is a model of the future Predicted values for each time step; This is the model prediction value after inverse normalization, which serves as the final prediction result. Although the above operations only use addition and multiplication transformations, the dual neural network structure can achieve complex linear or nonlinear mappings and adapt to the complexity of dynamically changing data.
[0088] S6: Compare the results with those of other optimization algorithms to determine the superiority of the proposed method in predicting thermal coal inventory.
[0089] Specifically, the evaluation indicators are shown in equation (7): (7) in, Indicates the sequence length. and These are the true value and the predicted value of the test set, respectively, used to compare the ratio of the model's prediction error to the true value. Their values range from 0 to positive infinity, with smaller values indicating better prediction performance.
[0090] Compared with the prior art, the advantages of the present invention are as follows: To address the characteristics of distribution drift and multi-scale coupling in time series data of thermal coal inventory, this embodiment employs an improved RevIn preprocessing algorithm that adaptively learns the normalization parameters of the historical window and the prediction window based on the dynamic patterns of the data. During the inverse normalization process, the prediction results are scaled to match the data distribution state of the prediction window. For the multi-scale coupling problem, the MTDNet model is used for prediction. This model separates the trend and seasonal terms of the downsampled series at each scale through a past information decomposition and mixing module, thereby isolating the impact of short-term fluctuations on long-term trends and capturing long-term fluctuation trends in the time series more precisely. Then, the seasonal and trend terms at different scales are mixed separately to allow information interaction at multiple scales, enhancing the model's ability to capture features at different time scales. A multi-prediction value mixing module uses multiple predictors to predict subsequences at each scale, and the results of these predictors are integrated to generate the final prediction value. This fully utilizes the complementarity of information at different scales, ultimately achieving better prediction accuracy for thermal coal inventory. Finally, prior knowledge terms of future distribution parameters are added to the model's loss function to reduce the impact of data noise on the optimized RevIn algorithm, resulting in better prediction accuracy.
[0091] Example 3: Another embodiment of this application relates to a device for predicting coal inventory for power generation. The implementation details of this device are described below. The following details are for ease of understanding and are not essential for implementing this solution. A schematic diagram of the coal inventory prediction device in this embodiment can be seen as follows: Figure 3 As shown, it includes an acquisition module 310, a normalization parameter generation module 320, a prediction module 330, and an inverse normalization module 340.
[0092] The acquisition module 310 is used to acquire multi-source time-series data related to thermal coal inventory; The normalization parameter generation module 320 is used to adaptively learn the normalization parameters of the historical window and the inverse normalization parameters of the prediction window based on the fluctuation pattern of the multi-source time series data. The prediction module 330 is used to perform normalization transformation on the multi-source time series data using the normalization parameters, and input the normalized data into a preset multi-scale time series prediction model to obtain a preliminary prediction result. The multi-scale time series prediction model is used to extract and fuse time series features of different time scales from the normalized data, and make predictions based on the fused multi-scale time series features to obtain the preliminary prediction result. The inverse normalization module 340 is used to inverse normalize the preliminary prediction results using the inverse normalization parameters to obtain the final predicted value of thermal coal inventory.
[0093] It is worth mentioning that all modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this application, this embodiment does not introduce units that are not closely related to solving the technical problems proposed in this application; however, this does not mean that other units are absent in this embodiment.
[0094] In some optional embodiments, the coal inventory forecasting device can implement the coal inventory forecasting method described in any of the above embodiments.
[0095] Example 4: Another embodiment of this application relates to an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the coal inventory forecasting method of the above embodiments.
[0096] In this embodiment, the memory and processor are connected via a bus, which can include any number of interconnected buses and bridges, connecting various circuits of one or more processors and the memory together. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be further described in this embodiment. A bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.
[0097] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.
[0098] Example 5: Another embodiment of this application relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the method embodiments described above.
[0099] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. 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.
[0100] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing this application, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of this application.
Claims
1. A method for predicting thermal coal inventory, characterized in that, include: Acquire multi-source time-series data related to thermal coal inventory; Based on the fluctuation patterns of the multi-source time series data, adaptive learning is used to obtain the normalized parameters of the historical window and the inverse normalized parameters of the prediction window. The normalization parameters are used to normalize the multi-source time series data, and the normalized data is input into a preset multi-scale time series prediction model to obtain preliminary prediction results. The multi-scale time series prediction model is used to extract and fuse time series features of different time scales from the normalized data, and to make predictions based on the fused multi-scale time series features to obtain the preliminary prediction results. The preliminary prediction results are inversely normalized using the inverse normalization parameter to obtain the final predicted value of thermal coal inventory.
2. The method for predicting thermal coal inventory according to claim 1, characterized in that, The adaptive learning of the fluctuation patterns of the multi-source time-series data to obtain normalized parameters for the historical window and inverse normalized parameters for the prediction window includes: By using dual neural networks, based on the same historical time series data, the normalized parameters of the historical window and the inverse normalized parameters of the prediction window are learned separately.
3. The method for predicting thermal coal inventory according to claim 1, characterized in that, The normalization transformation of the multi-source time series data is calculated according to the following formula: in, It is normalized data. It is historical time-series data. and These are affine parameters. and yes Time series The distribution of parameters in the input space describes the parameters.
4. The method for predicting thermal coal inventory according to claim 3, characterized in that, The preliminary prediction results are inversely normalized according to the following formula: in, The multi-scale time series prediction model is for the future The predicted value for each time step, i.e., the preliminary prediction result, This is the final predicted value of thermal coal inventory after denormalization. and yes Time series The parameters are distributed in the output space.
5. The method for predicting thermal coal inventory according to claim 1, characterized in that, Temporal features at different time scales are extracted and fused from the normalized data, including: The normalized data is decomposed into components representing periodic fluctuations and components representing long-term trends. For the periodic fluctuation component, a first-direction fusion operation is performed to fuse high-frequency scale information into low-frequency scale information; For the long-term trend component, a second-direction fusion operation is performed to fuse low-frequency scale information into high-frequency scale information. The periodic fluctuation components and the long-term trend components are added together after fusion to obtain the fused multi-scale time series features.
6. The method for predicting thermal coal inventory according to claim 1, characterized in that, Based on the fused multi-scale temporal features, a preliminary prediction result is obtained, including: Multiple parallel sub-predictors are used to predict the feature correspondences at different scales in the fused multi-scale temporal features respectively; The outputs of each of the sub-predictors are integrated to obtain the preliminary prediction result.
7. The method for predicting thermal coal inventory according to any one of claims 1-6, characterized in that, The loss function of the multi-scale time series prediction model includes a main loss term based on the error between the predicted and true values, and a regularization term used to constrain the learning effect of the inverse normalization parameters.
8. A device for predicting coal inventory for power generation, characterized in that, include: The acquisition module is used to acquire multi-source time-series data related to thermal coal inventory; The normalization parameter generation module is used to adaptively learn the normalization parameters of the historical window and the inverse normalization parameters of the prediction window based on the fluctuation patterns of the multi-source time series data. The prediction module is used to perform normalization transformation on the multi-source time series data using the normalization parameters, and input the normalized data into a preset multi-scale time series prediction model to obtain preliminary prediction results. The multi-scale time series prediction model is used to extract and fuse time series features of different time scales from the normalized data, and make predictions based on the fused multi-scale time series features to obtain the preliminary prediction results. The inverse normalization module is used to inverse normalize the preliminary prediction results using the inverse normalization parameters to obtain the final predicted value of thermal coal inventory.
9. An electronic device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method for predicting thermal coal inventory as described in any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the method for predicting thermal coal inventory as described in any one of claims 1 to 7.