Electricity coal inventory prediction method and device in coal-fired power generation scene, medium and equipment

Abstract

The embodiment of the invention provides an electricity-coal inventory prediction method and device in a coal-fired power generation scene, a medium and equipment. The method comprises the following steps: acquiring a feature data sequence of a target coal-fired power plant associated with electricity coal inventory in a first time window; wherein the first time window is a historical time window backtracking for a first preset number of days from the current date; the characteristic data sequence comprises daily historical coal storage amount, coal feed amount, coal consumption amount, power generation amount, temperature, freight rate and coal price in the first time window; inputting the feature data sequence into a preset electricity coal inventory prediction model, and outputting a corresponding daily electricity coal inventory prediction value in a second time window through the preset electricity coal inventory prediction model; wherein the second time window is a future time window of a second preset number of days before the current date. By using the method provided by the embodiment of the invention, high-precision and high-stability prediction of the electricity-coal inventory can be realized.

Description

Technical Field

[0001] This disclosure relates to the field of intelligent management technology for the power energy supply chain, specifically to a method, device, medium, and equipment for predicting coal inventory in a coal-fired power generation scenario. Background Technology

[0002] As a crucial pillar of my country's energy system, the stable supply of thermal coal is vital to the national economy and people's livelihoods. However, the geographical misalignment between major coal-producing areas and the eastern power load centers makes thermal coal inventory management a key aspect of ensuring the safe and economical operation of power plants. Accurate inventory forecasting can effectively balance supply security with inventory costs, which is crucial for the resilience of the energy supply chain.

[0003] Currently, forecasting methods in this field mainly rely on traditional time series models (such as ARIMA), machine learning algorithms (such as support vector machines and random forests), and deep learning models (such as LSTM and GRU). While these methods have achieved some success, they all have significant limitations. First, when faced with massive amounts of multi-dimensional features (such as production, transportation, meteorological, and power plant operation data), traditional methods and machine learning algorithms struggle to accurately select the most effective key features for forecasting, often relying on human experience or simple statistical indicators, resulting in poor performance. Second, coal inventory data contains complex multi-scale characteristics, including short-term intraday fluctuations and medium- to long-term seasonal and periodic patterns. Existing models are insufficient in effectively extracting and fusing these dynamic features across different time scales, leading to an incomplete understanding of the data's inherent patterns. Finally, time series data generally suffers from distribution drift, meaning the statistical characteristics of the data change over time, and existing models are poorly adapted to this, resulting in a significant reduction in their forecasting robustness and accuracy in dynamically changing environments. Summary of the Invention

[0004] To address the aforementioned technical problems, the present disclosure provides a solution. Embodiments of this disclosure offer a method, apparatus, medium, and equipment for predicting coal inventory in a coal-fired power generation scenario.

[0005] According to a first aspect of the present disclosure, a method for predicting coal inventory in a coal-fired power generation scenario is provided, wherein the method includes: Obtain the characteristic data sequence of the target coal-fired power plant associated with its coal inventory within a first time window; wherein, the first time window is a historical time window that traces back a first preset number of days from the current date; the characteristic data sequence includes the historical coal inventory, coal intake, coal consumption, power generation, temperature, freight rate, and coal price for each day within the first time window; The feature data sequence is input into a preset coal inventory prediction model, and the preset coal inventory prediction model outputs the corresponding daily coal inventory prediction value within the second time window. Wherein, the second time window is a future time window of a second preset number of days from the current date; the preset coal inventory prediction model is trained based on at least one set of training data pairs, and each set of training data pairs includes a feature data sample sequence and a corresponding coal inventory sample true value sequence.

[0006] According to a second aspect of the present disclosure, a device for predicting coal inventory in a coal-fired power generation scenario is provided, wherein the device includes: The data preprocessing unit is configured to: acquire a characteristic data sequence of the target coal-fired power plant associated with its coal inventory within a first time window; wherein, the first time window is a historical time window that traces back a first preset number of days from the current date; the characteristic data sequence includes the historical coal inventory, coal intake, coal consumption, power generation, temperature, freight rate, and coal price for each day within the first time window; The thermal coal inventory forecasting unit is configured to: input the feature data sequence into a preset thermal coal inventory forecasting model, and output the corresponding daily thermal coal inventory forecast value within a second time window via the preset thermal coal inventory forecasting model; Wherein, the second time window is a future time window of a second preset number of days from the current date; the preset coal inventory prediction model is trained based on at least one set of training data pairs, and each set of training data pairs includes a feature data sample sequence and a corresponding coal inventory sample true value sequence.

[0007] According to a third aspect of the present disclosure, an electronic device is provided, the electronic device comprising: a processor; a memory for storing executable instructions of the processor; the processor being configured to read the executable instructions from the memory and execute the instructions to implement the coal inventory prediction method for coal-fired power generation scenarios as described in the present disclosure.

[0008] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided, the storage medium storing a computer program for executing the coal inventory prediction method for a coal-fired power generation scenario as described in the present disclosure.

[0009] According to a fifth aspect of the present disclosure, a computer program product is provided, including a computer program, wherein when the computer program is executed by a processor, it implements the coal inventory prediction method for coal-fired power generation scenarios described in the present disclosure.

[0010] The coal inventory prediction method for coal-fired power generation scenarios disclosed herein ensures the criticality and effectiveness of the model input data through multi-stage joint feature screening, utilizes an improved RevIn algorithm to dynamically correct the distribution drift between training and prediction data to enhance model robustness, and combines an improved SCINet model to fully extract and integrate multi-scale time-series features of coal inventory changes to improve prediction accuracy. At the same time, by introducing prior knowledge of future distribution into the loss function of model training, the learning of distribution parameters is further optimized, thereby achieving highly accurate and stable prediction of coal inventory, providing reliable decision support for the scientific management and operation scheduling of power plant fuel. Attached Figure Description

[0011] The above and other objects, features, and advantages of this disclosure will become more apparent from the more detailed description of the embodiments thereof in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this disclosure and form part of the specification. They are used together with the embodiments of this disclosure to explain the disclosure and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0012] Figure 1 This is a flowchart illustrating a method for predicting coal inventory in a coal-fired power generation scenario provided by an exemplary embodiment of this disclosure. Figure 2 This is a public announcement Figure 1 An exemplary flowchart of a coal inventory prediction method for a coal-fired power generation scenario provided in the embodiments; Figure 3 This is a schematic diagram of the structure of a preset coal inventory forecasting model provided in an exemplary embodiment of this disclosure; Figure 4 This is a schematic diagram of the structure of a Back-Net unit provided in an exemplary embodiment of this disclosure; Figure 5 This is a schematic diagram of the structure of the SCINet unit and the basic module SCI-Block provided in an exemplary embodiment of this disclosure; Figure 6 This is a schematic diagram of the structure of the first learning transformation submodule provided in an exemplary embodiment of this disclosure; Figure 7 This is a public announcement Figure 1 Another exemplary flowchart of the coal inventory prediction method for coal-fired power generation scenario provided in the embodiments; Figure 8 This is a public announcement Figure 1 Another exemplary flowchart of the coal inventory prediction method for coal-fired power generation scenario provided in the embodiment; Figure 9This is a schematic diagram of the structure of a coal inventory prediction device for a coal-fired power generation scenario provided in an exemplary embodiment of this disclosure; Figure 10 This is a schematic diagram of the structure of an application embodiment of the electronic device disclosed herein. Detailed Implementation

[0013] The present disclosure will be further described below with reference to the embodiments shown in the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present disclosure, and not all embodiments of the present disclosure. It should be understood that the present disclosure is not limited to the exemplary embodiments described herein.

[0014] It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of this disclosure.

[0015] Those skilled in the art will understand that the terms "first," "second," etc., in the embodiments of this disclosure are only used to distinguish different steps, devices, or modules, and do not represent any specific technical meaning, nor do they indicate a necessary logical order between them.

[0016] It should also be understood that in the embodiments disclosed herein, "a plurality of" may refer to two or more, and "at least one" may refer to one, two or more.

[0017] It should also be understood that any component, data or structure mentioned in the embodiments of this disclosure can generally be understood as one or more unless expressly defined or given to the contrary in the context.

[0018] Furthermore, the term "and / or" in this disclosure is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this disclosure generally indicates that the preceding and following related objects have an "or" relationship.

[0019] It should also be understood that the description of the various embodiments in this disclosure emphasizes the differences between the various embodiments, and the similarities or similarities can be referred to each other. For the sake of brevity, they will not be described in detail.

[0020] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0021] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit this disclosure or its application or use.

[0022] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0023] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0024] Overview of the Invention Concept The core inventive concept of this disclosed technical solution lies in solving two major challenges in thermal coal inventory forecasting: data distribution drift and multi-scale feature coupling, through the synergistic effect of an improved RevIn algorithm and an improved SCINet model. Specifically, the RevIn algorithm, with its dual neural network structure, dynamically learns the data distribution parameters of historical and future windows, adaptively correcting distribution drift during normalization and denormalization. Simultaneously, the multi-scale decomposition and interactive feature extraction capabilities of the SCINet model effectively capture short-term fluctuations and long-term trends in inventory data. Furthermore, by introducing prior knowledge of future distribution into the loss function, the synergistic training of the two components is strengthened, ultimately achieving more accurate and robust forecasting of thermal coal inventory.

[0025] Based on the above-mentioned inventive concept, this disclosure proposes a coal inventory prediction scheme for coal-fired power generation scenarios as described in the following embodiments.

[0026] Example 1 Figure 1 This is a schematic diagram of a coal inventory prediction method for a coal-fired power generation scenario provided by an exemplary embodiment of this disclosure. The method can be executed on a server (e.g., a cloud service platform or a locally deployed server).

[0027] Specifically, refer to Figure 1 The method for predicting coal inventory in the context of coal-fired power generation includes: S110. Obtain the characteristic data sequence of the target coal-fired power plant associated with the coal inventory within the first time window.

[0028] The first time window is a historical time window that traces back a first preset number of days from the current date. The characteristic data sequence includes the historical coal inventory, coal intake, coal consumption, power generation, temperature, freight rates, and coal prices for each day within the first time window.

[0029] Optionally, refer to Figure 2 Step S110 can be implemented in the following way: S1110. Obtain the original feature data of the target coal-fired power plant related to the coal inventory within the first time window.

[0030] Specifically, historical data is collected from the target coal-fired power plant's business systems, sensors, and external data sources. The original characteristic data should cover the core supply chain links and external environmental factors affecting coal inventory, mainly including: historical coal inventory, coal intake, coal consumption, power generation, local temperature data, coal transportation prices, and coal market prices, among other multi-dimensional time series data.

[0031] S1120. Fill in missing values ​​and handle outliers in the original feature data to obtain preprocessed data.

[0032] Specifically, data interpolation is used to fill in missing values ​​generated during data acquisition to ensure the continuity of the time series. Simultaneously, the three-standard-deviation method is used to identify and remove outliers in the original feature data to eliminate noise interference. Furthermore, the statistical temporal resolution of all the above features is unified to the daily level, resulting in temporally aligned and complete preprocessed data.

[0033] S1130. Perform feature filtering on the preprocessed data to obtain the feature data sequence.

[0034] Specifically, a multi-stage joint screening strategy is employed to extract the most critical features for predicting thermal coal inventory from the preprocessed data, reconstructing a high-quality feature data sequence. First, random forest importance analysis is used for preliminary screening to assess the contribution of each feature to coal inventory prediction, retaining important features. Second, since the features initially screened may exhibit high collinearity, affecting model stability, correlation analysis is performed on the retained features to eliminate features highly similar to other feature data. Finally, to further ensure the causal relationship between the screened features and the prediction target, Granger causality tests are performed on the features after the aforementioned screening steps to verify whether their historical values ​​have significant predictive power for future coal inventory. Features that pass the test are ultimately retained, forming the feature data sequence used as model input.

[0035] S120. Input the feature data sequence into the preset coal inventory prediction model, and output the corresponding daily coal inventory prediction value within the second time window through the preset coal inventory prediction model.

[0036] Wherein, the second time window is a future time window of a second preset number of days from the current date; the preset coal inventory prediction model is trained based on at least one set of training data pairs, and each set of training data pairs includes a feature data sample sequence and a corresponding coal inventory sample true value sequence.

[0037] The preset coal inventory prediction model is a fusion model based on an improved RevIn algorithm and an improved SCINet model. The specific model structure is as follows: Optionally, refer to Figure 3 The preset coal inventory prediction model includes a cascaded normalized prediction network and a Forward-Net unit. The normalized prediction network includes a parallel first network branch and a second network branch; the first network branch is a data path; the second branch includes cascaded Back-Net units and SCINet units. The input of the data path and the input of the Back-Net unit together serve as the input of the normalized prediction network; the output of the data path is connected to the first input of the Forward-Net unit; and the output of the SCINet unit is connected to the second input of the Forward-Net unit.

[0038] The core of the aforementioned pre-designed thermal coal inventory forecasting model lies in a meticulously designed "reversible normalization-forecasting-denormalization" serial-parallel architecture. This architecture systematically solves key forecasting challenges through the collaborative work of two branches. Its workflow and principles can be described as follows: The original feature data sequence is first simultaneously input into the first and second network branches in parallel. The second network branch is the core processing path, cascading the Back-Net and SCINet units. The Back-Net unit performs the normalization step of the improved RevIn algorithm. As a neural network, it dynamically learns and extracts the distribution parameters (such as mean and variance) of historical input data and uses these parameters to adaptively normalize the original data. This step effectively eliminates the non-stationarity (i.e., distribution drift) in historical data, providing a more stable and learnable input representation for subsequent models. Next, the normalized data is fed into the SCINet unit, which is the core for multi-scale deep feature extraction and preliminary prediction. Through its unique sample convolution and interaction structure, it can simultaneously and effectively capture complex patterns at different time scales, such as short-term fluctuations and long-term trends, and outputs a preliminary prediction value located in a "stationary representation space."

[0039] Meanwhile, the first network branch (data path) directly transmits the raw data to the Forward-Net unit. This direct path preserves crucial contextual information about the overall distribution of the raw data for subsequent processes. Finally, the Forward-Net unit, acting as the denormalization step in the improved RevIn algorithm, receives the raw information from the data path and the preliminary predictions from the SCINet unit. Its core function is to learn and predict the data distribution parameters for future time windows, and then use these parameters to remap (i.e., denormalize) the preliminary predictions output by SCINet, which are in a stationary space, to a distribution that conforms to the future evolution trend of the actual data, thereby generating the final, high-precision forecast of thermal coal inventory.

[0040] In summary, this model corrects distribution drift in time-series data through a dynamic, reversible normalization framework consisting of Back-Net and Forward-Net. Within this framework, SCINet units focus on extracting deep, multi-scale features from stationary data. The entire architecture forms a synergistically optimized whole, enabling the final prediction results to not only profoundly reflect the inherent complex patterns of the data but also accurately align with future data distribution trends.

[0041] Optionally, refer to Figure 4 The Back-Net unit includes a cascaded input layer, multiple hidden modules, and an output layer; wherein each hidden module includes a fully connected layer and an activation function.

[0042] The Forward-Net unit includes a cascaded input layer, multiple hidden modules, and an output layer; each hidden module includes a fully connected layer and an activation function. That is, both the Back-Net unit and the Forward-Net unit can be fully connected neural networks.

[0043] As attached Figure 4 As shown, both Back-Net and Forward-Net units consist of an input layer, multiple cascaded hidden modules, and an output layer. Each hidden module performs a "fully connected layer + activation function" operation. This structural choice is based on the characteristics of the task. The core task of Back-Net and Forward-Net is to learn a mapping relationship: mapping the input time-series data (or its representation) to key parameters describing the characteristics of that data distribution. Fully connected neural networks are powerful and general-purpose function approximators, capable of automatically extracting high-level features from the input and fitting complex mapping functions through multiple layers of nonlinear transformations, making them very suitable for this parameter fitting task. The cascading of multiple hidden modules provides sufficient network depth and nonlinearity to capture the potential nonlinear, high-dimensional correlation between the data distribution and the input features.

[0044] Taking the Back-Net unit as an example, its workflow can be symbolically described as follows: given a sequence of feature data for a historical window. As input, the data is first received and shaped by the input layer. Then, it passes through N hidden modules sequentially. In each module, the data first undergoes linear transformation and feature combination via a fully connected layer, which can be expressed as: Then, nonlinearity is introduced through an activation function (such as the ReLU function), expressed as: This combination of linear transformation and nonlinear activation is repeated across multiple modules, enabling the network to abstract and learn deep patterns in the data layer by layer. Finally, the output layer (usually a fully connected layer) transforms the output of the last hidden module into a parameter vector of the target dimension. For Back-Net, its output is the learned historical window data distribution parameters, which can be denoted as the mean. and variance Similarly, the structure of the Forward-Net unit is exactly the same. It takes the preliminary prediction results from the previous stage and the contextual information of the original data as input, goes through the same "fully connected + activation" layer-by-layer processing, and finally outputs the predicted future window data distribution parameters. and .

[0045] Adopting such Figure 4 The fully connected neural network structure shown is used to construct the Back-Net and Forward-Net units, providing a stable and learnable distribution parameter estimator for the entire prediction model. The Back-Net parameterizes the statistics used for normalization through its network parameterization, while the Forward-Net parameterizes the statistics used for inverse normalization. Together, they achieve the learnability and adaptability of the reversible instance normalization (RevIn) process.

[0046] Additionally, it should be noted that: As attached Figure 4 As shown, although the Back-Net and Forward-Net cells have the same structure, their functions and computational objectives are different, and their computational processes can be clearly expressed by formulas respectively.

[0047] The Back-Net unit uses the aforementioned feature data sequence (historical window data). As input, the sequence undergoes nonlinear transformation and feature abstraction sequentially through the input layer and multiple fully connected + activated hidden modules. The final output is assumed to be the learned parameters of the historical window data distribution, i.e., the scale parameter. (Related to variance) and displacement parameters (Related to the mean). The calculation process for this unit can be abstracted as follows:

[0048] in, This represents all trainable network parameters for the Back-Net unit. These parameters are then used to perform a learnable adaptive normalization of the input data (i.e., an improved RevIn process), as shown in the following formula:

[0049] Where H is the normalized, stationary data representation fed into subsequent SCINet units for prediction. and These are learnable affine transformation parameters used to enhance the expressive power of the model.

[0050] Similarly, the Forward-Net unit outputs the initial prediction results from the SCINet unit. (Located in a stable representation space) and necessary contextual information (usually including some aggregate representation of the original data) are taken as input, and processed through its own "input layer-hidden module-output layer" structure. Its goal is to learn the data distribution parameters for predicting future windows. and The calculation process can be abstracted as follows:

[0051] in, This indicates a splicing operation; the context can come from the original information in the data path. These are the trainable parameters of the Forward-Net units. Finally, these predicted future distribution parameters are used to inversely normalize the initial output of SCINet to obtain the final predicted value. :

[0052] This formula is the inverse process of the aforementioned normalization formula; it measures the predicted values ​​in the stationary space. Remapping to a distribution that aligns with future data expectations , On a certain scale.

[0053] As mentioned above, Figure 4 The fully connected neural network structure shown is the specific carrier for the Back-Net and Forward-Net units to implement the aforementioned parameter learning function. Through end-to-end training, they learn to accurately estimate the distribution parameters, thereby driving the entire reversible normalization framework to operate effectively and dynamically adapt to changes in data distribution.

[0054] Optionally, refer to Figure 5The SCINet unit includes cascaded binary tree subnetworks, a fusion and splicing layer, and a fully connected layer; wherein, the input end of the binary tree subnetwork and the output end of the fusion and splicing layer are connected to realize element-wise addition operation; the binary tree subnetwork includes multiple basic modules SCI-Block cascaded in the form of binary trees.

[0055] As attached Figure 5 As shown in (b), the SCINet unit is a hierarchical architecture. Its core is a binary tree subnetwork composed of multiple (L=3 in the figure) SCI-Block basic modules cascaded in a binary tree form. The normalized sequence output from the previous stage (such as the Back-Net unit) is used as input. Each level of SCI-Block processes the input sequence, and its output is integrated and rearranged through a fusion and splicing layer. Finally, it is mapped to the output features of the current level through a fully connected layer. The bridging connection (implied by the plus sign before "Concat&Realign" in the figure) directly adds the input of the binary tree subnetwork to the output of the fusion and splicing layer element-wise. This residual connection structure helps to alleviate gradient vanishing and promotes deep network training. This multi-level cascaded structure enables the model to recursively downsample and interactively learn the sequence, thereby gradually extracting and fusing features from different time scales, from fine to coarse.

[0056] Optionally, refer to Figure 5 The basic module SCI-Block includes cascaded segmentation layers and an odd-even interaction subnetwork. The odd-even interaction subnetwork includes parallel odd paths, even paths, and a first learning transformation submodule connecting the odd and even paths. Second Learning Transformation Submodule The third learning transformation submodule The fourth learning transformation submodule The odd-numbered paths include cascaded first Hadamard product operation nodes and element-wise addition operation nodes; wherein the output of the first Hadamard product operation node is connected to the first input of the element-wise addition operation node; the even-numbered paths include cascaded second Hadamard product operation nodes and element-wise subtraction operation nodes; wherein the output of the second Hadamard product operation node is connected to the first input of the element-wise subtraction operation node. The first output of the segmentation layer is connected to the first input of the first Hadamard product operation node and the input of the first learning transformation submodule, respectively; the second output of the segmentation layer is connected to the first input of the second Hadamard product operation node and the input of the second learning transformation submodule, respectively. The output of the first learning transformation submodule is connected to the second input of the second Hadamard product operation node; the output of the second learning transformation submodule is connected to the second input of the first Hadamard product operation node. The output of the first Hadamard product operation node is also connected to the input of the third learning transformation submodule; the output of the second Hadamard product operation node is also connected to the input of the fourth learning transformation submodule. The output of the third learning transformation submodule is connected to the second input of the element-by-element subtraction operation node; the output of the fourth learning transformation submodule is connected to the second input of the element-by-element addition operation node. The output of the element-by-element subtraction operation node serves as the output of the even-numbered path; the output of the element-by-element addition operation node serves as the output of the odd-numbered path.

[0057] like Figure 5 As shown in (a), SCI-Block is the basic unit for constructing SCINet. Its core principle is to achieve feature downsampling and information fusion through parity-even interaction. First, the split layer divides the input feature sequence F into two sub-sequences according to the parity of the time step index: F odd (Odd Index) and F even (Even index). These two subsequences are then processed in depth within the odd-even interaction subnetwork.

[0058] This subnetwork contains parallel odd-numbered and even-numbered paths, as well as four key learning transformation submodules (corresponding to the symbols in the diagram). These submodules are typically composed of lightweight neural networks (such as multilayer perceptrons, MLPs) used to learn complex interactions between features. The specific interaction process is as follows: On odd-numbered paths, F odd On the one hand, it comes from even-numbered paths, via submodules The transformed information undergoes a first Hadamard product (element-by-element multiplication) to obtain an intermediate result; on the other hand, this intermediate result is then processed by a submodule. Further transformations are performed. Finally, this intermediate result is also combined with data from even-numbered paths, via submodules. The transformed information is then fused using element-wise addition to output a new odd-numbered path representation. .

[0059] On even-numbered paths, F even On the one hand, it comes from odd-numbered paths, via submodules The transformed information undergoes a second Hadamard product (element-by-element multiplication) to obtain an intermediate result; on the other hand, this intermediate result is then processed by a submodule. Further transformations are performed. Ultimately, this intermediate result is also combined with data from odd-numbered paths, via submodules. The transformed information is then fused using element-wise subtraction to output a new odd-numbered path representation. .

[0060] Through the interaction of the above design and Each component incorporates information from another path, enabling feature interaction and refinement across time steps, while the sequence length is downsampled to half of its original length. It usually carries refined high-frequency (detailed) information, while This contains more low-frequency (trend) information.

[0061] Optionally, refer to Figure 6 The first, second, third, and fourth learning transformation submodules have the same internal hierarchical structure. The first learning transformation submodule includes a sequentially cascaded copy padding layer, a one-dimensional encoding convolutional layer, a LeakyReLU function, a random dropout layer, a one-dimensional decoding convolutional layer, and a Tanh function.

[0062] like Figure 6 As shown, these four key sub-modules share the same internal structure, each consisting of a series of sequentially cascaded neural network layers, forming a lightweight encoder-decoder one-dimensional convolutional feature transformer. Its core principle is to non-linearly compress, transform, and reconstruct the input feature segments to learn complex time dependencies and interaction patterns across odd and even sequences. This structure is specifically designed for processing time series data, and the specific data processing flow is as follows: The input feature sequence X (with shape (C, L), where C is the feature dimension and L is the sequence length) first enters the ReplicationPad1d layer. This step pads the sequence by copying edge values ​​at both ends, aiming to ensure that subsequent convolutional operations do not change the temporal length L of the sequence, maintaining temporal alignment, which is crucial for interactive operations that require precise temporal correspondence.

[0063] The padded data is then passed through a one-dimensional encoding convolutional layer (Conv1d). This layer is configured with C input channels, h×C output channels (where h is a spread factor greater than 1), and a kernel size of k. Its function is to downsample and encode the input features, expanding the feature dimension (from C to hC) and utilizing the kernel k to sense local time windows, thereby extracting richer, higher-dimensional temporal abstractions. Subsequently, a LeakyReLU activation function (with a negative slope of slope=0.01) introduces non-linearity into the network, allowing small negative values ​​to pass through, which helps alleviate the vanishing gradient problem and enhances the model's expressive power.

[0064] Subsequently, the network uses a dropout layer with a high dropout probability (e.g., p=0.5). This layer randomly "shuts down" a portion of neurons during training, which is an effective regularization technique to prevent the model from overfitting to specific noise patterns in the training data and improve the generalization ability of the learned interaction transformation rules.

[0065] The regularized features are then processed by a one-dimensional decoding convolutional layer (Conv1d). This layer compresses the feature dimension from the high-dimensional hC back to the original C, achieving feature reconstruction and dimensionality reduction. Finally, the Tanh activation function constrains the output value to [ Within the range of [1,1]. This bounded output is very important because it ensures that when subsequent Hadamard products (element-wise multiplication) or addition / subtraction interactions are performed with the original path, the transformed result is a bounded, controllable "modulation signal" or "residual correction", thus stabilizing the training process.

[0066] It also needs to be explained that, Figure 6 The defined symmetric structure of "encoding-nonlinear activation-regularization-decoding" is the core of the learnable function in SCI-Block that enables effective feature interactions. Four sub-modules. Although their functional roles are slightly different (each responsible for information filtering and transformation between different paths), they share the same essential task (feature transformation). Therefore, this unified and efficient lightweight convolutional structure is adopted to control model complexity while ensuring powerful representation capabilities.

[0067] Example 2 Based on the above embodiment 1, as an optional implementation method, refer to Figure 7 The method further includes the step of training the preset thermal coal inventory prediction model: S210. Obtain the at least one set of training data pairs.

[0068] Understandably, at least one set of training data pairs is constructed from historical data. Each data pair includes a feature data sample sequence and its corresponding ground truth sequence of thermal coal inventory samples. The cutoff date of the third time window (i.e., the historical observation window for model input) covered by the feature data sample sequence should be earlier than the start date of the prediction time window corresponding to the ground truth sequence of thermal coal inventory samples, to ensure that the training conforms to the actual forward-looking prediction scenario. Specifically, the feature data sample sequence contains multi-dimensional features for each day within the third time window, such as historical coal inventory, coal intake, coal consumption, power generation, temperature, freight rates, and coal prices. The ground truth sequence of thermal coal inventory samples contains the actual daily thermal coal inventory within the prediction time window, serving as the target value for model learning.

[0069] S220. Input the feature data sample sequence from each training data pair into the coal inventory prediction model to be trained, and output the corresponding coal inventory prediction value sequence through the coal inventory prediction model to be trained.

[0070] The coal inventory prediction model to be trained is the model of the preset coal inventory prediction model without training and parameter adjustment.

[0071] Specifically, the feature data sample sequences from each training data pair are input into an untrained (parameters uninitialized or in a random initial state) coal inventory prediction model to be trained. The structure of this model to be trained is exactly the same as the preset coal inventory prediction model described in this embodiment (i.e., the model including the improved RevIn and the improved SCINet). The model processes the input sequence, sequentially performing normalization by the Back-Net unit, multi-scale feature extraction and prediction by the SCINet unit, and denormalization by the Forward-Net unit, finally outputting the corresponding coal inventory prediction value sequence.

[0072] S230. Using a preset loss function, based on the true value sequence of thermal coal inventory samples and the corresponding predicted value sequence of thermal coal inventory in each training data pair, obtain the function value of the preset loss function.

[0073] As an optional example, S230 can use the following formula to calculate the function value of the preset loss function. ;

[0074] Wherein, K represents the number of feature data sample sequences used in one training iteration; N represents the total number of feature data sample sequences based on the at least one set of training data pairs; This represents the prior term weight hyperparameter; This indicates the length of the prediction time window; iIndex representing the sequence of feature data samples; Indicates the corresponding number i Feature data sample sequence, from date arrive Forecast values ​​for thermal coal inventory; Indicates the corresponding number i Feature data sample sequence, from date arrive True values ​​of thermal coal inventory samples; Indicates the corresponding number i A sequence of characteristic data samples, starting from date arrive The average of the true values ​​of the thermal coal inventory sample; Indicates the corresponding number i A sequence of characteristic data samples, on the date The distribution parameters are predicted by the Forward-Net cells.

[0075] The design principle of this loss function is as follows: The first term is the standard mean squared error (MSE), which directly measures the overall deviation between the predicted sequence and the true sequence. The second term is an introduced prior knowledge term, which calculates the difference between the average inventory of the true future window and the mean of the distribution predicted by Forward-Net. The purpose of adding this term is to provide an explicit supervision signal to the learning process of the Forward-Net units, forcing its predicted distribution parameters to be as close as possible to the true statistical characteristics of the future data, thereby improving the model's adaptability to data distribution drift and the stability of the prediction. Finally, by substituting the predicted results and true values ​​of each data pair into the above formula, the total loss value under the current model parameters can be calculated.

[0076] S240. Based on the function value of the preset loss function, train the coal inventory prediction model to be trained until the preset training completion condition is met, and obtain the preset coal inventory prediction model from the coal inventory prediction model to be trained.

[0077] As an optional example, see [reference] Figure 8 Step S240 can be implemented in the following way: S2410. The function value of the preset loss function is passed back to the coal inventory prediction model to be trained.

[0078] Specifically, the loss value Backpropagation is performed back to the coal inventory prediction model to be trained, and the gradient of the loss function with respect to the parameters of each layer of the model (including all trainable parameters in Back-Net, SCINet, and Forward-Net) is calculated.

[0079] S2420. The coal inventory prediction model to be trained adjusts the network parameters of each network layer according to the function value of the preset loss function.

[0080] Specifically, optimization algorithms (such as Adam) are used to adjust (update) all weights and bias parameters of each network layer in the model based on the calculated gradient direction, so as to reduce the loss function value.

[0081] S2430. Iteratively execute the steps from obtaining the function value of the preset loss function to adjusting the network parameters of each network layer until the preset training completion condition is met.

[0082] Specifically, steps S220 to S2420 are repeated, i.e., multiple iterations of training are performed. In each iteration, using new or resampled batches of training data, the processes of forward prediction, loss calculation, backpropagation, and parameter update are repeated until the model performance converges on the validation set or reaches the preset number of iterations (i.e., the preset training completion condition is met). The model obtained at this point is the trained and usable preset coal inventory prediction model.

[0083] The coal inventory prediction method for coal-fired power generation scenarios provided in Embodiments 1 and 2 of this disclosure ensures the criticality and effectiveness of the model input data through multi-stage joint feature screening, utilizes the improved RevIn algorithm to dynamically correct the distribution drift between training data and prediction data to enhance model robustness, and combines the improved SCINet model to fully extract and integrate multi-scale time-series features of coal inventory changes to improve prediction accuracy. At the same time, by introducing prior knowledge of future distribution into the loss function of model training, the learning of distribution parameters is further optimized, thereby achieving highly accurate and stable prediction of coal inventory, providing reliable decision support for the scientific management and operation scheduling of power plant fuel.

[0084] Example 3 It should be understood that the coal inventory forecasting method for coal-fired power generation scenarios described in the foregoing embodiments of this document can also be similarly applied to the coal inventory forecasting devices for the following coal-fired power generation scenarios for similar extension. For simplicity, these are not described in detail.

[0085] Figure 9 This is a schematic diagram of a coal inventory prediction device for a coal-fired power generation scenario provided in an exemplary embodiment of this disclosure. (Refer to...) Figure 9 The device includes: The data preprocessing unit 110 is configured to: acquire a characteristic data sequence of the target coal-fired power plant associated with its coal inventory within a first time window; wherein, the first time window is a historical time window that traces back a first preset number of days from the current date; the characteristic data sequence includes the historical coal inventory, coal intake, coal consumption, power generation, temperature, freight rate, and coal price for each day within the first time window; The thermal coal inventory forecasting unit 120 is configured to: input the feature data sequence into a preset thermal coal inventory forecasting model, and output the corresponding daily thermal coal inventory forecast value within a second time window via the preset thermal coal inventory forecasting model. Wherein, the second time window is a future time window of a second preset number of days from the current date; the preset coal inventory prediction model is trained based on at least one set of training data pairs, and each set of training data pairs includes a feature data sample sequence and a corresponding coal inventory sample true value sequence.

[0086] The coal inventory prediction device for coal-fired power generation scenarios provided in the above embodiments of this disclosure ensures the criticality and effectiveness of the model input data through multi-stage joint feature screening, uses an improved RevIn algorithm to dynamically correct the distribution drift between training data and prediction data to enhance model robustness, and combines an improved SCINet model to fully extract and integrate multi-scale time-series features of coal inventory changes to improve prediction accuracy. At the same time, by introducing prior knowledge of future distribution into the loss function of model training, the learning of distribution parameters is further optimized, thereby achieving highly accurate and stable prediction of coal inventory, providing reliable decision support for the scientific management and operation scheduling of power plant fuel.

[0087] Example 4 In addition, this disclosure also provides an electronic device, including: a memory for storing a computer program; and a processor for executing the computer program stored in the memory, wherein when the computer program is executed, it implements the coal inventory prediction method for coal-fired power generation scenarios described in any of the above embodiments of this disclosure.

[0088] Figure 10 This is a schematic diagram of the structure of an application embodiment of the electronic device disclosed herein. Below, reference is made to… Figure 10 This describes an electronic device according to embodiments of the present disclosure. The electronic device may be either or both of a first device and a second device, or a standalone device independent of them, which may communicate with the first device and the second device to receive acquired input signals from them.

[0089] like Figure 10As shown, the electronic device includes one or more processors and memory. The processor may be a central processing unit (CPU) or other processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device to perform desired functions. The memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor may execute the program instructions to implement the coal inventory prediction method for coal-fired power generation scenarios in the various embodiments of this disclosure described above, and / or other desired functions.

[0090] In one example, the electronic device may further include input and output devices, which are interconnected via a bus system and / or other forms of connection mechanisms (not shown). Furthermore, the input device may include, for example, a keyboard, a mouse, etc. The output device can output various information to the outside, including determined distance information, direction information, etc. The output device may include, for example, a display, a speaker, a printer, and a communication network and its connected remote output devices, etc.

[0091] Of course, for the sake of simplicity, Figure 10 Only some of the components of the electronic device relevant to this disclosure are shown, omitting components such as buses, input / output interfaces, etc. In addition, the electronic device may include any other suitable components depending on the specific application.

[0092] In addition to the methods and apparatus described above, embodiments of this disclosure may also be computer program products, including computer program instructions that, when executed by a processor, cause the processor to perform the steps in the coal inventory prediction method for coal-fired power generation scenarios according to various embodiments of this disclosure as described in the foregoing portion of this specification.

[0093] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this disclosure. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on a user's computing device, partially on a user's computing device, as a standalone software package, partially on a user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0094] Furthermore, embodiments of this disclosure may also be computer-readable storage media storing computer program instructions thereon, which, when executed by a processor, cause the processor to perform the steps in the coal inventory prediction method for coal-fired power generation scenarios according to various embodiments of this disclosure as described in the foregoing portion of this specification.

[0095] The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0096] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as ROM, RAM, magnetic disk, or optical disk.

[0097] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0098] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For system embodiments, since they largely correspond to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0099] The block diagrams of devices, apparatuses, devices, and systems disclosed herein are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0100] The methods and apparatus of this disclosure may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the methods is for illustrative purposes only, and the steps of the methods of this disclosure are not limited to the order specifically described above unless otherwise specifically stated. Furthermore, in some embodiments, this disclosure may also be implemented as a program recorded on a recording medium, the program including machine-readable instructions for implementing the methods according to this disclosure. Thus, this disclosure also covers recording media storing programs for performing the methods according to this disclosure.

[0101] It should also be noted that in the apparatus, devices, and methods of this disclosure, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions to this disclosure.

[0102] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0103] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A method for predicting coal inventory in a coal-fired power generation scenario, characterized in that, The method includes: Obtain the characteristic data sequence of the target coal-fired power plant associated with its coal inventory within a first time window; wherein, the first time window is a historical time window that traces back a first preset number of days from the current date; the characteristic data sequence includes the historical coal inventory, coal intake, coal consumption, power generation, temperature, freight rate, and coal price for each day within the first time window; The feature data sequence is input into a preset coal inventory prediction model, and the preset coal inventory prediction model outputs the corresponding daily coal inventory prediction value within the second time window. Wherein, the second time window is a future time window of a second preset number of days from the current date; the preset coal inventory prediction model is trained based on at least one set of training data pairs, and each set of training data pairs includes a feature data sample sequence and a corresponding coal inventory sample true value sequence.

2. The method according to claim 1, characterized in that, The acquisition of the characteristic data sequence associated with the target coal-fired power plant and its coal inventory within the first time window includes: Obtain the original characteristic data of the target coal-fired power plant related to the coal inventory within the first time window; The original feature data is filled with missing values ​​and outliers are processed to obtain preprocessed data; The preprocessed data is subjected to feature filtering to obtain the feature data sequence.

3. The method according to claim 1, characterized in that, The preset coal inventory prediction model includes a cascaded normalized prediction network and a Forward-Net unit; The normalized prediction network includes a first network branch and a second network branch in parallel. The first network branch is a data path; The second branch includes cascaded Back-Net units and SCINet units; in, The input end of the data path and the input end of the Back-Net unit together serve as the input end of the normalized prediction network; The output of the data path is connected to the first input of the Forward-Net unit; The output of the SCINet unit is connected to the second input of the Forward-Net unit.

4. The method according to claim 3, characterized in that, The Back-Net unit includes a cascaded input layer, multiple hidden modules, and an output layer; wherein each hidden module includes a fully connected layer and an activation function. The Forward-Net unit includes a cascaded input layer, multiple hidden modules, and an output layer; wherein each hidden module includes a fully connected layer and an activation function.

5. The method according to claim 3, characterized in that, The SCINet unit includes a cascaded binary tree subnetwork, a fusion and splicing layer, and a fully connected layer; wherein, the input end of the binary tree subnetwork and the output end of the fusion and splicing layer are connected to realize element-wise addition operation; The binary tree subnetwork includes multiple basic modules SCI-Block that are cascaded in the form of a binary tree.

6. The method according to claim 5, characterized in that, The basic module SCI-Block includes cascaded segmentation layers and odd-even interaction subnetworks; The odd-even interaction subnetwork includes parallel odd paths, even paths, and a first learning transformation submodule, a second learning transformation submodule, a third learning transformation submodule, and a fourth learning transformation submodule connected between the odd paths and even paths. The odd path includes a cascaded first Hadamard product operation node and an element-wise addition operation node; wherein the output of the first Hadamard product operation node is connected to the first input of the element-wise addition operation node; The even-numbered path includes a cascaded second Hadamard product operation node and an element-by-element subtraction operation node; wherein the output of the second Hadamard product operation node is connected to the first input of the element-by-element subtraction operation node. The first output of the segmentation layer is connected to the first input of the first Hadamard product operation node and the input of the first learning transformation submodule, respectively. The second output of the segmentation layer is connected to the first input of the second Hadamard product operation node and the input of the second learning transformation submodule, respectively. The output of the first learning transformation submodule is connected to the second input of the second Hadamard product operation node; The output of the second learning transformation submodule is connected to the second input of the first Hadamard product operation node; The output of the first Hadamard product operation node is also connected to the input of the third learning transformation submodule; The output of the second Hadamard product operation node is also connected to the input of the fourth learning transformation submodule; The output of the third learning transformation submodule is connected to the second input of the element-by-element subtraction operation node; The output of the fourth learning transformation submodule is connected to the second input of the element-by-element addition operation node; The output of the element-by-element subtraction operation node serves as the output of the even-numbered path. The output of the element-wise addition operation node serves as the output of the odd-numbered path.

7. The method according to claim 1, characterized in that, The method also includes the step of training the preset thermal coal inventory prediction model: Obtain the at least one set of training data pairs; Wherein, the end date of the third time window corresponding to the feature data sample sequence is later than the start date of the prediction time window corresponding to the true value sequence of the thermal coal inventory sample; The feature data sample sequence includes the daily historical coal inventory, coal intake, coal consumption, power generation, temperature, freight rate, and coal price within the third time window. The true value sequence of thermal coal inventory samples includes the true values ​​of thermal coal inventory samples for each day within the prediction time window. The feature data sample sequence from each training data pair is input into the coal inventory prediction model to be trained, and the coal inventory prediction model to be trained outputs the corresponding coal inventory prediction value sequence; wherein, the coal inventory prediction model to be trained is the model of the preset coal inventory prediction model without training and parameter tuning; Using a preset loss function, the function value of the preset loss function is obtained based on the true value sequence of thermal coal inventory samples and the corresponding predicted value sequence of thermal coal inventory in each training data pair; Based on the function value of the preset loss function, the coal inventory prediction model to be trained is trained until the preset training completion condition is met, and the preset coal inventory prediction model is obtained from the coal inventory prediction model to be trained.

8. A coal inventory prediction device for coal-fired power generation, characterized in that, The device includes: The data preprocessing unit is configured to: acquire a characteristic data sequence of the target coal-fired power plant associated with its coal inventory within a first time window; wherein, the first time window is a historical time window that traces back a first preset number of days from the current date; the characteristic data sequence includes the historical coal inventory, coal intake, coal consumption, power generation, temperature, freight rate, and coal price for each day within the first time window; The thermal coal inventory forecasting unit is configured to: input the feature data sequence into a preset thermal coal inventory forecasting model, and output the corresponding daily thermal coal inventory forecast value within a second time window via the preset thermal coal inventory forecasting model; Wherein, the second time window is a future time window of a second preset number of days from the current date; the preset coal inventory prediction model is trained based on at least one set of training data pairs, and each set of training data pairs includes a feature data sample sequence and a corresponding coal inventory sample true value sequence.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which is used to execute the coal inventory prediction method for coal-fired power generation scenarios as described in claims 1 to 7.

10. An electronic device, characterized in that, The electronic device includes: a processor; a memory for storing executable instructions of the processor; the processor is configured to read the executable instructions from the memory and execute the instructions to implement the coal inventory prediction method for coal-fired power generation scenarios as described in claims 1 to 7.

Images