Power plant coal consumption prediction method, device, equipment, medium and product

Abstract

The invention relates to the technical field of power plant coal consumption prediction, in particular to a power plant coal consumption prediction method and device, equipment, a medium and a product, and the method comprises the steps: collecting air temperature data and power generation data corresponding to the coal consumption of a power plant, and obtaining a power plant coal consumption data set; dividing the power plant coal consumption data set into a multivariable time sequence, a future covariable sequence and a moment target sequence, and setting a resolution parameter; building a covariable condition diffusion model oriented to time multi-resolution power plant coal consumption prediction; a historical time sequence and a future covariable sequence in the multivariable time sequence are transmitted into a conditional channel of the covariable conditional diffusion model, a moment target sequence is transmitted into a target channel of the covariable conditional diffusion model, and the coal consumption of the power plant is predicted; the power plant coal consumption prediction capability can be improved.

Description

Technical Field

[0001] This invention relates to the field of power plant coal consumption prediction technology, and particularly to a method, apparatus, equipment, medium, and product for predicting power plant coal consumption. Background Technology

[0002] The evolution of coal consumption in power plants is typically highly uncertain, influenced by factors such as fluctuations in industrial load and sudden temperature changes. This uncertainty stems from the nonlinearity of the system itself and disturbances caused by external covariate factors like weather conditions. While many related technologies employ point-based prediction strategies, these methods struggle to characterize multi-peak distributions and extreme fluctuations, resulting in significant deficiencies in the reliability and stability of the predictions. Therefore, there is a technical challenge in this field to improve the ability to predict coal consumption in power plants. Summary of the Invention

[0003] This invention provides a method, apparatus, equipment, medium, and product for predicting coal consumption in power plants, solving the technical problem of how to improve the coal consumption prediction capability of power plants.

[0004] In a first aspect, the present invention provides a method for predicting coal consumption in power plants, comprising: step S01, collecting temperature data and power generation data corresponding to the coal consumption of power plants to obtain a coal consumption dataset; step S02, dividing the coal consumption dataset into a multivariate time series, a future covariate series, and a time target series, and setting resolution parameters; step S03, constructing a covariate conditional diffusion model for predicting coal consumption in power plants at multiple time resolutions; and step S04, inputting the historical time series and the future covariate series from the multivariate time series into the conditional channel of the covariate conditional diffusion model, and inputting the time target series into the target channel of the covariate conditional diffusion model to predict the coal consumption of power plants.

[0005] In some embodiments, step S03, the step of building a covariate conditional diffusion model for predicting coal consumption in power plants over time multi-resolution, includes: step S31, adding Gaussian noise to a given original target sequence to construct at least one noisy sample; step S32, performing conditional modeling on the noisy sample based on an inverse denoising network to recover the original target sequence.

[0006] In some embodiments, the inverse denoising network in step S32 is composed of multiple layers of residual blocks stacked together. Each residual block includes: a diffusion step embedding module, a conditional information fusion module, and a spatiotemporal Transformer module. The diffusion step embedding module encodes the current diffusion step number into a vector through position encoding, and then passes it through a fully connected layer and SiLU activation to obtain the embedding representation of the diffusion step number. The conditional information fusion module maps the resolution category embedding and the temporal step embedding representation to the same dimension as the main network channel, and expands it to the same length as the input sequence before injecting it into the residual block through addition. The spatiotemporal Transformer module captures the long-range dependencies in the sequence time through the temporal layer and models the interaction structure between multiple variables through the feature layer. The output of the spatiotemporal Transformer module is fed into the residual connection branch and the skip connection branch, and the skip connection results are accumulated layer by layer to obtain the noise prediction.

[0007] In some embodiments, the spatiotemporal Transformer module includes: a time layer for capturing long-range dependencies of a sequence in time; a feature layer for modeling the interaction structure among multiple variables; and co-modeling the outputs of the time layer and the feature layer to obtain the time series modeling result.

[0008] In some embodiments, step S04, which involves inputting the historical time series and future covariate series from the multivariate time series into the conditional channel of the covariate conditional diffusion model, includes: step S41, inputting the historical time series into the conditional channel, wherein the historical time series reflects past trends and cycles; step S42, inputting the future covariate series into the conditional channel, wherein the future covariate series provides known information about the future; and step S43, inputting the resolution category embedding to control the time granularity of the prediction output.

[0009] In some embodiments, step S43, the step of inputting resolution category embedding, includes: step S431, assigning a learnable resolution category embedding vector to each temporal resolution category; step S432, adding the resolution category embedding vector to the temporal step embedding representation obtained by position encoding of the temporal step to obtain the fused conditional embedding representation; step S433, inputting the fused conditional embedding representation into each layer of the diffusion denoising network as a global control signal.

[0010] Secondly, the present invention provides a power plant coal consumption prediction device, comprising: a data collection module for collecting temperature data and power generation data corresponding to the power plant coal consumption to obtain a power plant coal consumption dataset; a data partitioning module for partitioning the power plant coal consumption dataset into a multivariate time series, a future covariate series, and a time target series, and setting resolution parameters; a model building module for building a covariate conditional diffusion model for predicting power plant coal consumption at multiple time resolutions; and a prediction module for inputting the historical time series and the future covariate series from the multivariate time series into the conditional channel of the covariate conditional diffusion model, and inputting the time target series into the target channel of the covariate conditional diffusion model to predict the power plant coal consumption.

[0011] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method of any of the above aspects.

[0012] Fourthly, the present invention provides an electronic device including a processor and a memory, wherein a computer program is stored in the memory, and the processor executes the computer program to implement the method of any of the above aspects.

[0013] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements a method for any of the above aspects.

[0014] This invention provides a method, apparatus, equipment, medium, and product for predicting coal consumption in power plants. The method includes: collecting temperature data and power generation data corresponding to the coal consumption of power plants to obtain a coal consumption dataset; dividing the coal consumption dataset into a multivariate time series, a future covariate series, and a time-target series, and setting resolution parameters; constructing a covariate conditional diffusion model for predicting coal consumption in power plants at multiple time resolutions; inputting the historical time series and the future covariate series from the multivariate time series into the conditional channel of the covariate conditional diffusion model, and inputting the time-target series into the target channel of the covariate conditional diffusion model to predict the coal consumption of power plants; thereby improving the coal consumption prediction capability of power plants. Attached Figure Description

[0015] The invention will now be described in more detail with reference to embodiments and the accompanying drawings: Figure 1 This is a schematic diagram of a power plant coal consumption prediction method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a power plant coal consumption prediction device provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a denoising network model structure provided in an embodiment of the present invention; Figure 4This is a schematic diagram of a multi-resolution covariate conditional diffusion model architecture provided in an embodiment of the present invention.

[0016] In the accompanying drawings, the same parts are referred to by the same reference numerals, and the drawings are not drawn to scale. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present invention and to fully understand and implement the process of how the present invention uses technical means to solve technical problems and achieve corresponding technical effects, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. The embodiments of the present invention and the various features therein can be combined with each other without conflict, and the resulting technical solutions are all within the protection scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0018] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0019] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0020] The evolution of coal consumption in power plants is typically highly uncertain, influenced by factors such as fluctuations in industrial load and sudden temperature changes. This uncertainty stems from the nonlinearity of the system itself and disturbances caused by external covariate factors such as weather conditions. While many related technologies employ point prediction strategies, these methods struggle to characterize scenarios involving multi-peak distributions and extreme fluctuations, resulting in significant deficiencies in the reliability and stability of the prediction results.

[0021] There is a technical problem in this field of how to improve the ability to predict coal consumption in power plants.

[0022] To address the aforementioned technical problem of how to improve the coal consumption prediction capability of power plants, this invention proposes a method, apparatus, equipment, medium, and product for predicting coal consumption in power plants. The implementation details of this invention are described below. The following content is only for ease of understanding and is not essential for implementing this solution.

[0023] Example 1 Figure 1 This is a flowchart illustrating a power plant coal consumption prediction method provided in an embodiment of this application, as shown below. Figure 1 As shown, in this embodiment, a method for predicting power plant coal consumption is provided, including: step S01, collecting temperature data and power generation data corresponding to power plant coal consumption to obtain a power plant coal consumption dataset; step S02, dividing the power plant coal consumption dataset into a multivariate time series, a future covariate series, and a time target series, and setting resolution parameters; step S03, building a covariate conditional diffusion model for predicting power plant coal consumption at multiple time resolutions; and step S04, inputting the historical time series and future covariate series from the multivariate time series into the conditional channel of the covariate conditional diffusion model, and inputting the time target series into the target channel of the covariate conditional diffusion model to predict power plant coal consumption.

[0024] In actual power plant operation, dispatchers require coal consumption forecasts at different time granularities: for example, short-term dispatch requires a 1-day granularity, while medium-term planning requires a 3-day or 7-day granularity. Traditional methods require training a separate model for each granularity, which not only increases training costs but also easily leads to inconsistent forecast results at different granularities. Furthermore, traditional point forecasts cannot characterize the multi-peak distribution caused by extreme weather (such as sudden temperature changes) or load fluctuations, making it difficult to meet the technical requirements of scenarios such as risk warning. The technical problem to be solved in this embodiment is how to support multiple time granularity tasks for power plant coal consumption forecasting under the same model structure, while improving the model's adaptability to future coal consumption trends and its ability to describe uncertain structures.

[0025] In this embodiment, the technical solution first collects temperature data and power generation data corresponding to the coal consumption of power plants to obtain a dataset. Then, the dataset is divided into multivariate time series (including historical coal consumption trends), future covariate series (such as future weather forecasts), and time-target series (the coal consumption value to be predicted), and resolution parameters are set (such as 1 day, 3 days, etc.). Next, a time-multi-resolution covariate conditional diffusion model is built, which can generate a probabilistic generation model of diverse future possibility distributions. Finally, the historical time series and future covariate series are input into the model's conditional channel to provide contextual information for prediction, and the time-target series is input into the target channel to predict the coal consumption of power plants. Here, covariates refer to additional information that improves prediction performance, and resolution parameters are used to control the time granularity of the prediction output.

[0026] The technical solution in this embodiment, through a unified covariate conditional diffusion model structure, enables the same model to support coal consumption prediction tasks at multiple time granularities, eliminating the need to train a separate model for each granularity and reducing training and maintenance costs. Simultaneously, the diffusion model's generation mechanism outputs the future distribution of coal consumption rather than a single value, characterizing potential technical scenarios such as extreme fluctuations and improving the reliability and stability of the prediction. Furthermore, the conditional channel integrates historical trends with future covariates (such as temperature), allowing the model to better adapt to future trend changes. For example, when future covariates indicate a sudden drop in temperature, the model can generate multiple possible distributions of rising coal consumption.

[0027] Example 2 Based on the above embodiments, step S03, the step of building a covariate conditional diffusion model for predicting coal consumption in power plants at multiple time resolutions, includes: step S31, adding Gaussian noise to the given original target sequence to construct at least one noisy sample; step S32, performing conditional modeling on the noisy sample based on an inverse denoising network to recover the original target sequence.

[0028] In power plant coal consumption forecasting, generative models need to generate future sequences that conform to actual coal consumption patterns from random noise. However, traditional generative methods struggle to control the constraints of the generation process (such as incorporating future temperature covariates) and cannot achieve a gradual reconstruction from noise to a valid sequence, resulting in low-quality generated sequences that are out of touch with reality. This embodiment aims to address the technical problem of constructing a covariate conditional diffusion model to achieve a gradual reconstruction of the power plant's future coal consumption forecast sequence from pure noise.

[0029] In this embodiment, the construction of the covariate conditional diffusion model is divided into two stages: a diffusion stage and a generation stage. In the diffusion stage, Gaussian noise is progressively added to the given original power plant coal consumption target sequence to construct multiple noisy samples, gradually transforming the original sequence into a noise distribution. In the generation stage, a reverse denoising network is used to conditionally model the noisy samples, recovering the original target sequence. The reverse denoising network is the core of the model; it learns to predict noise from the noisy samples and progressively denoise them, using historical coal consumption sequences and future covariates (such as weather forecasts) as conditions, ultimately generating a future coal consumption sequence that meets the conditional constraints.

[0030] The technical solution of this embodiment achieves a gradual reconstruction from pure noise to a power plant coal consumption sequence through a two-stage design of diffusion and generation: the diffusion stage provides a clear probabilistic path for the generation process, and the generation stage controls the generation direction by combining a reverse denoising network with conditional information (such as future temperature) to ensure that the generated sequence conforms to the actual coal consumption pattern; for example, in the generation stage, the model can gradually adjust the denoising process based on the future high temperature covariate to generate a reasonable sequence of rising coal consumption; in addition, this process can generate diverse future possibilities, providing power plants with a more comprehensive risk assessment basis. Compared with traditional generative methods, the generated sequence has higher quality and stronger adaptability to actual scenarios.

[0031] Example 3 Based on the above embodiments, the inverse denoising network in step S32 is composed of multiple layers of residual blocks stacked together. Each residual block includes: a diffusion step embedding module, a conditional information fusion module, and a spatiotemporal Transformer module. The diffusion step embedding module obtains a vector by position encoding the current diffusion step number, and then passes it through a fully connected layer and SiLU activation to obtain the embedding representation of the diffusion step number. The conditional information fusion module maps the resolution category embedding and the time step embedding representation to the same dimension as the main network channel, and expands it to the same length as the input sequence before injecting it into the residual block through addition. The spatiotemporal Transformer module captures the long-range dependencies in the sequence time through the time layer and models the interaction structure between multiple variables through the feature layer. The output of the spatiotemporal Transformer module is fed into the residual connection branch and the skip connection branch, and the skip connection results are accumulated layer by layer to obtain the noise prediction.

[0032] The technical problem this embodiment aims to solve is how to achieve high-quality conditional generation modeling and enhance the ability of the inverse denoising network to understand and model complex dynamic patterns of power plant coal consumption sequences. Power plant coal consumption sequences are influenced by multiple factors, such as long-term trends in historical coal consumption, short-term fluctuations in future temperatures, and periodic characteristics at different time granularities. Traditional denoising networks struggle to effectively integrate this conditional information and cannot simultaneously capture long-term temporal dependencies (such as seasonal variations) and multivariate interactions (such as the synergistic effect of temperature and power generation), resulting in low denoising accuracy and generated sequences that do not conform to actual patterns.

[0033] In this embodiment, the inverse denoising network is composed of multiple stacked residual blocks. Each residual block includes: a diffusion step embedding module, a conditional information fusion module, and a spatiotemporal Transformer module. The diffusion step embedding module converts the current diffusion step number into a vector through position encoding, and then obtains an embedded representation through a fully connected layer and SiLU activation, which is used to mark the current denoising step. The conditional information fusion module maps the resolution category embedding (controlling the prediction granularity) and time step embedding to the main network channel dimension, expands it to the sequence length, and injects it into the residual block to achieve conditional regulation. The spatiotemporal Transformer module captures the long-range dependence of the coal consumption sequence (such as seasonal changes) through the time layer and models multivariate interactions (such as the relationship between temperature and power generation) through the feature layer. Finally, the output of the residual block is sent to the residual connection and skip connection branches, and the skip connection results are accumulated layer by layer to obtain the noise prediction.

[0034] The technical solution in this embodiment ensures the stability of deep network training through the stacking design of multi-layer residual blocks; the diffusion step embedding allows the model to perceive the current denoising stage; the conditional information fusion module effectively integrates resolution and time step information, enabling the model to adapt to multi-granularity tasks; the spatiotemporal Transformer module simultaneously captures long-term time dependencies and multivariate interactions, such as recognizing the long-term trend of higher coal consumption in winter than in summer, and the multivariate relationship of increased power generation leading to increased coal consumption when temperatures drop sharply; these designs collectively enhance the conditional generation capability of the denoising network, resulting in power plant coal consumption sequences that are more consistent with actual patterns and significantly improving denoising accuracy.

[0035] Example 4 Based on the above embodiments, the spatiotemporal Transformer module includes: a time layer for capturing long-range dependencies of sequences in time; a feature layer for modeling the interaction structure between multiple variables; and the outputs of the time layer and the feature layer are used to co-model to obtain the time series modeling result.

[0036] The technical problem this embodiment aims to solve is how to effectively capture the long-term temporal dependencies and multivariate interaction structures of power plant coal consumption sequences, thereby improving the model's ability to model complex dynamic patterns. Power plant coal consumption sequences exhibit significant long-term temporal dependencies, such as the periodicity of peak coal consumption each winter; simultaneously, they are influenced by multivariate interactions, such as increased heating load due to lower temperatures, which in turn increases power generation and coal consumption. Traditional models struggle to efficiently capture both of these relationships simultaneously, leading to predictions that ignore key influencing factors and suffer from insufficient accuracy.

[0037] In this embodiment, the spatiotemporal Transformer module of the inverse denoising network consists of two sub-layers: a time layer and a feature layer. The time layer is used to capture the long-range dependencies of the coal consumption sequence over time: for example, through an attention mechanism, the model can associate the current coal consumption value with the peak winter coal consumption data from several months ago to identify periodic trends. The feature layer is used to model the interaction structure between multiple variables: for example, analyzing the relationship between temperature, power generation, and coal consumption; when the temperature decreases, the model can capture the synchronous change of increased power generation and rising coal consumption. The two sub-layers work together to model the complex dynamic patterns of the coal consumption sequence and output a feature representation that integrates time dependence and multivariate interaction information.

[0038] The technical solution in this embodiment effectively solves the problem that traditional models struggle to balance long-term time dependence and multivariate interactions through the design of two sub-layers in the spatiotemporal Transformer module: the time layer enables the model to accurately identify the long-term cyclical characteristics of coal consumption, such as the upward trend in coal consumption brought about by winter heating; the feature layer enables the model to understand the synergistic impact of multiple factors on coal consumption, for example, when future covariates show a sharp drop in temperature and a planned increase in power generation, the model can predict that coal consumption will increase significantly; these capabilities improve the model's accuracy in modeling complex coal consumption patterns, and the generated prediction sequences are more in line with actual operating scenarios.

[0039] Example 5 Based on the above embodiments, step S04, which involves inputting the historical time series and future covariate series from the multivariate time series into the conditional channel of the covariate conditional diffusion model, includes: step S41, inputting the historical time series into the conditional channel, wherein the historical time series reflects past trends and cycles; step S42, inputting the future covariate series into the conditional channel, wherein the future covariate series provides known information about the future; and step S43, inputting the resolution category embedding to control the time granularity of the prediction output.

[0040] The technical problem this embodiment aims to solve is how to improve the distribution modeling capability, multi-task adaptability, and generalization capability of power plant coal consumption prediction models by integrating multiple types of conditional information. In power plant coal consumption prediction, historical coal consumption sequences reflect past trends, while future covariates (such as weather forecasts) provide forward-looking information, and resolution category embedding controls the prediction granularity. Traditional models often use only single-conditional information or cannot effectively integrate multiple types of conditions, resulting in models that are difficult to adapt to different task scenarios, have poor generalization ability, and cannot accurately characterize the uncertainty distribution of coal consumption.

[0041] In this embodiment, the historical time series and future covariate series from the multivariate time series are input into the conditional channel of the covariate conditional diffusion model, along with resolution category embedding. The historical time series reflects the power plant's past coal consumption trends and cyclical characteristics; for example, daily coal consumption data from the past six months can demonstrate a weekly cycle. The future covariate series provides known information about the future, such as temperature forecasts for the next 7 days and holiday markings. The resolution category embedding controls the time granularity of the prediction output, such as 1-day, 3-day, or 7-day granularity. These three types of information work together on the model's denoising network, allowing the model to consider historical trends, future influencing factors, and target granularity simultaneously during the generation process, achieving high-quality conditional generation modeling.

[0042] The technical solution in this embodiment significantly improves the overall performance of the model by integrating three types of key conditional information: historical time series enable the model to capture the basic trend of coal consumption, future covariates enable the model to be forward-looking (e.g., predicting coal consumption increases by combining future high temperatures), and resolution category embedding enables the model to support multi-granularity tasks. The synergistic effect of these conditional information not only enhances the model's distributed modeling ability and generates diverse future possibilities for coal consumption, but also improves the model's multi-task adaptability and generalization ability. For example, the same model can simultaneously meet the needs of short-term scheduling (1-day granularity) and medium-term planning (7-day granularity), and the prediction results remain consistent across different granularities, providing power plants with more flexible and reliable decision support.

[0043] Example 6 Based on the above embodiments, step S43, the step of inputting resolution category embedding, includes: step S431, assigning a learnable resolution category embedding vector to each temporal resolution category; step S432, adding the resolution category embedding vector to the temporal step embedding representation obtained by position encoding of the temporal step to obtain the fused conditional embedding representation; step S433, inputting the fused conditional embedding representation into each layer of the diffusion denoising network as a global control signal.

[0044] The technical problem this embodiment aims to solve is how to inject time-granularity information into the diffusion model to achieve unified modeling of power plant coal consumption prediction tasks with different time resolutions, thereby improving the model's generalization ability and robustness across time-granularity tasks. In power plant operation, different scenarios require coal consumption prediction at different granularities: for example, daily scheduling uses a 1-day granularity, while monthly planning uses a 7-day granularity. Traditional methods require training the model separately for each granularity, which not only increases training costs but also easily leads to inconsistent prediction results at different granularities. Furthermore, the model has difficulty transferring knowledge (such as from a 1-day granularity to a 7-day granularity), resulting in poor generalization ability.

[0045] In this embodiment, the step of inputting the resolution category embedding includes: first, assigning a learnable embedding vector to each time resolution category, such as vector A for 1-day granularity and vector B for 3-day granularity; then, adding the resolution category embedding to the time step embedding representation to obtain the fused conditional embedding representation; finally, inputting the fused embedding representation into each layer of the diffusion denoising network as a global control signal. Here, the resolution category embedding is a vector representation of discrete granularity labels, and the global control signal guides the model to adapt to the coal consumption sequence features of the corresponding granularity during the generation process; for example, a 7-day granularity needs to capture trends over a longer period.

[0046] The technical solution of this embodiment, through a resolution category embedding mechanism, enables the same model to support coal consumption prediction tasks at multiple time granularities: the model does not need to be trained separately for each granularity, reducing training and maintenance costs; the learnability of the embedding vectors allows the model to automatically learn the feature differences of different granularities, for example, the 1-day granularity focuses on short-term load fluctuations, while the 7-day granularity focuses on long-term trends; in addition, the knowledge transfer capability is improved, for example, after the model learns the relationship between coal consumption and temperature at the 1-day granularity, it can quickly adapt to the prediction task at the 7-day granularity; thus, the model exhibits stronger generalization ability and robustness in cross-time granularity tasks, and the prediction results remain consistent across different granularities.

[0047] Example 7 Figure 2 This is a schematic diagram of the structure of a power plant coal consumption prediction device provided in an embodiment of this application, as shown below. Figure 2 As shown in the technical solution of this embodiment, a power plant coal consumption prediction device is provided, including: a data collection module, used to collect temperature data and power generation data corresponding to the power plant coal consumption to obtain a power plant coal consumption dataset; a data partitioning module, used to partition the power plant coal consumption dataset into a multivariate time series, a future covariate series, and a time target series, and set resolution parameters; a model building module, used to build a covariate conditional diffusion model for time-multi-resolution power plant coal consumption prediction; and a prediction module, used to input the historical time series and future covariate series from the multivariate time series into the conditional channel of the covariate conditional diffusion model, and input the time target series into the target channel of the covariate conditional diffusion model to predict the power plant coal consumption.

[0048] In actual power plant operation, dispatchers require coal consumption forecasts at different time granularities: for example, short-term dispatch requires a 1-day granularity, while medium-term planning requires a 3-day or 7-day granularity. Traditional methods require training a separate model for each granularity, which not only increases training costs but also easily leads to inconsistent forecast results at different granularities. Furthermore, traditional point forecasts cannot characterize the multi-peak distribution caused by extreme weather (such as sudden temperature changes) or load fluctuations, making it difficult to meet the technical requirements of scenarios such as risk warning. The technical problem to be solved in this embodiment is how to support multiple time granularity tasks for power plant coal consumption forecasting under the same model structure, while improving the model's adaptability to future coal consumption trends and its ability to describe uncertain structures.

[0049] In this embodiment, the technical solution first collects temperature data and power generation data corresponding to the coal consumption of power plants to obtain a dataset. Then, the dataset is divided into multivariate time series (including historical coal consumption trends), future covariate series (such as future weather forecasts), and time-target series (the coal consumption value to be predicted), and resolution parameters are set (such as 1 day, 3 days, etc.). Next, a time-multi-resolution covariate conditional diffusion model is built, which can generate a probabilistic generation model of diverse future possibility distributions. Finally, the historical time series and future covariate series are input into the model's conditional channel to provide contextual information for prediction, and the time-target series is input into the target channel to predict the coal consumption of power plants. Here, covariates refer to additional information that improves prediction performance, and resolution parameters are used to control the time granularity of the prediction output.

[0050] The technical solution in this embodiment, through a unified covariate conditional diffusion model structure, enables the same model to support coal consumption prediction tasks at multiple time granularities, eliminating the need to train a separate model for each granularity and reducing training and maintenance costs. Simultaneously, the diffusion model's generation mechanism outputs the future distribution of coal consumption rather than a single value, characterizing potential technical scenarios such as extreme fluctuations and improving the reliability and stability of the prediction. Furthermore, the conditional channel integrates historical trends with future covariates (such as temperature), allowing the model to better adapt to future trend changes. For example, when future covariates indicate a sudden drop in temperature, the model can generate multiple possible distributions of rising coal consumption.

[0051] Other technical features and effects of this embodiment correspond to those of the above embodiments, and will not be repeated here.

[0052] Example 8 In this embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, it implements the method of any one of the above embodiments, including: A method for predicting coal consumption in power plants includes: Step S01, collecting temperature data and power generation data corresponding to the coal consumption of power plants to obtain a coal consumption dataset; Step S02, dividing the coal consumption dataset into a multivariate time series, a future covariate series, and a time target series, and setting resolution parameters; Step S03, building a covariate conditional diffusion model for predicting coal consumption in power plants at multiple time resolutions; Step S04, inputting the historical time series and the future covariate series from the multivariate time series into the conditional channel of the covariate conditional diffusion model, and inputting the time target series into the target channel of the covariate conditional diffusion model to predict the coal consumption of power plants.

[0053] In some embodiments, step S03, the step of building a covariate conditional diffusion model for predicting coal consumption in power plants over time multi-resolution, includes: step S31, adding Gaussian noise to a given original target sequence to construct at least one noisy sample; step S32, performing conditional modeling on the noisy sample based on an inverse denoising network to recover the original target sequence.

[0054] In some embodiments, the inverse denoising network in step S32 is composed of multiple layers of residual blocks stacked together. Each residual block includes: a diffusion step embedding module, a conditional information fusion module, and a spatiotemporal Transformer module. The diffusion step embedding module encodes the current diffusion step number into a vector through position encoding, and then passes it through a fully connected layer and SiLU activation to obtain the embedding representation of the diffusion step number. The conditional information fusion module maps the resolution category embedding and the temporal step embedding representation to the same dimension as the main network channel, and expands it to the same length as the input sequence before injecting it into the residual block through addition. The spatiotemporal Transformer module captures the long-range dependencies in the sequence time through the temporal layer and models the interaction structure between multiple variables through the feature layer. The output of the spatiotemporal Transformer module is fed into the residual connection branch and the skip connection branch, and the skip connection results are accumulated layer by layer to obtain the noise prediction.

[0055] In some embodiments, the spatiotemporal Transformer module includes: a time layer for capturing long-range dependencies of a sequence in time; a feature layer for modeling the interaction structure among multiple variables; and co-modeling the outputs of the time layer and the feature layer to obtain the time series modeling result.

[0056] In some embodiments, step S04, which involves inputting the historical time series and future covariate series from the multivariate time series into the conditional channel of the covariate conditional diffusion model, includes: step S41, inputting the historical time series into the conditional channel, wherein the historical time series reflects past trends and cycles; step S42, inputting the future covariate series into the conditional channel, wherein the future covariate series provides known information about the future; and step S43, inputting the resolution category embedding to control the time granularity of the prediction output.

[0057] In some embodiments, step S43, the step of inputting resolution category embedding, includes: step S431, assigning a learnable resolution category embedding vector to each temporal resolution category; step S432, adding the resolution category embedding vector to the temporal step embedding representation obtained by position encoding of the temporal step to obtain the fused conditional embedding representation; step S433, inputting the fused conditional embedding representation into each layer of the diffusion denoising network as a global control signal.

[0058] In the technical solution of this embodiment, an electronic device is provided, including a processor and a memory. The memory stores a computer program, and the processor executes the computer program to implement the method of any of the above embodiments.

[0059] In the technical solution of this embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the method of any of the above embodiments.

[0060] The processor may include, but is not limited to, one or more processors or microprocessors. Each processor may be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic component, for performing the methods in the above embodiments. The computer-readable storage medium may be implemented by any type of volatile or non-volatile storage device or a combination thereof, and may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, EPROM memory, EEPROM memory, registers, computer storage media (e.g., hard disk, floppy disk, solid-state drive, removable disk, CD-ROM, DVD-ROM, Blu-ray disc, etc.).

[0061] Computer-readable storage media may also store at least one computer-executable program / instruction, such as computer-readable instructions. Computer-readable storage media include, but are not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Computer-readable storage media may include, for example, read-only memory (ROM), hard disk, flash memory, etc. For example, a non-transitory computer-readable storage medium may be connected to a computing device such as a computer, and then, when the computing device executes the computer-readable instructions stored on the computer-readable storage medium, the various methods described above can be performed.

[0062] In addition, the computer device may also include (but is not limited to) a data bus, an input / output (I / O) bus, a display, and input / output devices (e.g., a keyboard, mouse, speakers, etc.). The processor can communicate with external devices via the I / O bus through a wired or wireless network. In one embodiment, the at least one computer-executable instruction may also be compiled into or comprise a software product / computer program product, wherein one or more computer-executable instructions, when executed by the processor, perform the steps of the various functions and / or methods in the embodiments described herein.

[0063] Example 9 Based on the above embodiments, this embodiment provides an application example.

[0064] This application example belongs to the field of time series forecasting and is a covariate conditional diffusion model method for multi-resolution power plant coal consumption forecasting.

[0065] In practical power plant coal consumption forecasting tasks, the future evolution of power plant coal consumption is typically highly uncertain, due to factors such as fluctuations in industrial load start-up and shutdown, and sudden temperature changes. This uncertainty stems from the nonlinearity of the system itself, disturbances from external covariate factors such as weather, and dynamic changes in user demand structure. Traditional methods often employ point forecasting strategies, which involve learning a mapping function to output a single estimate of the future power plant coal consumption sequence. However, point forecasting methods struggle to characterize potential future scenarios such as multi-peak distributions and extreme fluctuations, resulting in significant deficiencies in the reliability and stability of the forecast results, particularly in critical scenarios requiring advance scheduling, peak shaving, or risk warnings.

[0066] To overcome the aforementioned limitations, this invention proposes a covariate conditional diffusion model for multi-resolution power plant coal consumption prediction, serving as a unified extension to existing deterministic and generative methods. This method uses historical time series and future covariates as conditional inputs, employing a diffusion model's generative framework to gradually reconstruct the future prediction sequence from pure noise. Simultaneously, by introducing resolution category embeddings, the model can support prediction tasks at multiple time granularities within the same network structure, achieving flexible and consistent multi-resolution modeling. This design not only significantly improves the model's adaptability to future trends but also enhances its ability to describe uncertain structures.

[0067] Currently, with the development of deep learning, a large number of methods relying on deterministic functions to model future values ​​have emerged in the field of time series prediction, becoming one of the mainstream solutions. These methods typically take historical observation sequences as input and directly output predicted values ​​for several future moments through deep neural network structures. TimeXer uses a token-mixing architecture to replace the traditional attention structure, improving modeling efficiency and training stability. TiDE employs a time decoupling mechanism to achieve independent modeling of long-term trends and short-term disturbances, enhancing multi-step prediction capabilities. SCINet, based on a hierarchical decomposition strategy, separates signals from a temporal perspective, improving interpretability and modeling granularity. iTransformer introduces linear attention and temporal position embedding structures to improve the ability to capture high-frequency features. PatchTST, drawing inspiration from image patch modeling, divides the time series into local patches and extracts long-term dependencies by stacking Transformer layers, achieving leading performance on multiple benchmark datasets. These models achieve good results in short- and medium-term prediction tasks by learning the mapping relationship between fixed-length historical windows and future windows. Their implementation often uses a sequence encoder to extract dynamic features of the time series, and then a predictor outputs the future sequence.

[0068] While these methods achieve high prediction accuracy at specific resolutions, they generally suffer from two limitations: first, they lack the ability to describe prediction uncertainty, failing to generate diverse predictions or estimate confidence intervals; second, they typically require separate model training for tasks with different time resolutions, lacking a unified approach to handling multi-granularity prediction problems. These issues limit their generalization ability and practicality in complex power plant coal consumption prediction scenarios.

[0069] The purpose of this solution is to enable the power plant coal consumption prediction model to have uncertainty modeling capabilities, covariate condition modeling capabilities, and to solve the resolution generalization problem.

[0070] The purpose of this invention is to address the shortcomings of related technologies, such as poor ability to describe uncertainty and low prediction accuracy, and to propose a covariate conditional diffusion model for multi-resolution power plant coal consumption prediction.

[0071] The purpose of this invention is to address the shortcomings of existing algorithms, such as poor ability to describe uncertainty and low prediction accuracy, and to propose a covariate conditional diffusion model for multi-resolution power plant coal consumption prediction. To achieve the above objective, this invention employs the following technical solution: S1: Collect and organize data related to the supply chain network, temperature, and power generation of power plants; S2: Divide the dataset into a multivariate time series of length and a sequence of covariates known in the future step, and set the resolution parameter; S3: A time-multi-resolution covariate conditional diffusion model is built using a conditional diffusion generation framework; S4: Input the historical coal inventory multivariate time series and the known future covariate series into the condition channel, and input the time target series into the target channel; S5: Using a time-multi-resolution covariate conditional diffusion model to predict various possible future coal consumption rates for power plants; S6: The final prediction result is obtained by using the output of the jump connection normalization model; S7: Compare the results with those of other optimization algorithms to determine the superiority of the proposed method in predicting coal consumption in power plants.

[0072] Figure 3 This is a schematic diagram of a denoising network model structure provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a multi-resolution covariate conditional diffusion model architecture provided in an embodiment of the present invention.

[0073] The conditional diffusion generation framework consists of two main stages: (1) the diffusion stage, which uses the forward process setting of the standard diffusion model to construct a series of intermediate variables as noisy samples by gradually adding Gaussian noise to the given target sequence; (2) the generation stage, which requires the model to learn the inverse denoising network to conditionally model the noisy samples, thereby gradually recovering the original target sequence.

[0074] Three types of key information are introduced as input conditions: (1) historical time series, reflecting past trends and cycles; (2) future covariate series, providing known information about the future; and (3) resolution category embedding, controlling the time granularity of the prediction output. The above information works together on the denoising network of the diffusion model, enabling the model to not only have a strong distribution modeling capability, but also to flexibly cope with multi-task settings, improving the generalization ability and uncertainty expression capability of the prediction.

[0075] Among them, high-quality conditional generation modeling uses the Transformer residual convolutional network as a parameterized form of the denoising function to progressively predict noise in the diffusion process. It is composed of multiple layers of stacked residual blocks and has both temporal modeling and conditional control capabilities.

[0076] The model input is the noisy target sequence at the current diffusion step, including two input channels: a conditional channel and a target channel, representing the historical / covariate input and the predicted target variable, respectively. The initial input is processed by a 1×1 convolution with a ReLU activation function for channel mapping, and then fed into stacked residual layers for temporal modeling.

[0077] Each residual module includes the following key components: diffusion step embedding module, conditional information fusion module, and spatiotemporal Transformer module.

[0078] The diffusion step embedding module encodes the current step number into a 128-dimensional vector through position encoding. This vector is then passed through a fully connected layer and SiLU activation to obtain the embedding representation of the current diffusion step number, which is then input into subsequent modules.

[0079] The conditional information fusion module is used to encode additional conditional information, such as temporal resolution category information. This invention uses a 2-layer MLP to map the resolution category embedding and time step embedding representation to the same dimension as the main network channel. After extending the result to the same length as the input sequence, it is injected into each residual block through addition, thereby achieving dynamic control of the entire diffusion process.

[0080] The representation after conditional injection is input into the spatiotemporal Transformer module, which consists of two sub-layers: a temporal layer and a feature layer. The former captures the long-range dependencies of the sequence in time, while the latter models the interaction structure among multiple variables. The two work together to significantly enhance the understanding and modeling capabilities of complex dynamic patterns.

[0081] Finally, the output of the residual block is fed into the residual connection branch and the skip connection branch, respectively. The former is used to maintain stable training of the deeper layers of the network, while the latter is used to aggregate the intermediate features of all layers as the basis for the final output. The skip connection results of all residual layers are accumulated layer by layer, and then mapped to the target channel dimension through two layers of 1×1 convolution and ReLU activation to output the noise prediction.

[0082] The generation process can be flexibly controlled through conditional input, including the following two conditional mechanisms: (1) Future covariate modeling mechanism Future covariates It includes known external information about the target time period, such as weather forecasts, holiday markers, control variables, and event-driven signals. In time series forecasting, covariates often have a direct guiding effect on future trends, and how to fully model their temporal structure and dynamic impact is crucial to improving forecast quality.

[0083] In this model, covariate modeling employs a concatenated conditional coding strategy, which involves concatenating historical time series and future covariates along the time dimension to form a unified conditional input sequence, which serves as the conditional channel input for the model. Specifically, let the historical observation sequence be... The future covariate sequence is First, it is mapped to a feature space of the same dimension using two independent linear projection modules: (1) Then and By splicing the data along the time dimension, a total length of [length missing] is formed. Condition sequence: (2) This splicing sequence serves as a conditional channel. The input is mapped to the number of channels through convolution after entering the network, and then used as explicit conditional information in the attention mechanism and regulation module in the subsequent residual blocks.

[0084] This concatenation strategy allows the model to simultaneously receive past trends and known future external drivers, resulting in a more forward-looking generative bias. During the modeling process, the network can selectively focus on the entire conditional sequence using an attention mechanism, not only capturing long-term historical dependencies but also identifying potential inflection points and prior signals in covariates, thereby improving its responsiveness to future time series trends.

[0085] (2) Resolution category embedding mechanism To support multi-resolution prediction tasks, this invention introduces an explicit resolution control mechanism, which injects the temporal granularity information corresponding to the current task into the diffusion model through category embedding.

[0086] set up This represents all supported prediction granularity categories (e.g., 1 day, 3 days, 7 days, etc.), each category can be viewed as a discrete label. This invention assigns a learnable embedding vector to each resolution category: (3) Resolution embedding and time step After positional encoding, the resulting embedded representations are summed and input into each layer of the diffusion denoising network as global control signals. Different resolution tasks share model parameters, distinguishing prediction targets only through class embeddings, which facilitates transfer generalization. For example, after learning the generation pattern of "3-day prediction," the model can transfer some structural knowledge to "7-day prediction," improving sample efficiency. This mechanism not only enables the model to adapt to different resolutions but also improves cross-task consistency, allowing the model to exhibit stronger generalization ability and robustness when facing tasks with diverse prediction granularities.

[0087] This method is based on the generation mechanism of the diffusion model. It constructs a complex distribution by progressively adding and denoising the target sequence, and introduces three types of key information as input conditions: (1) historical power plant coal consumption sequence, reflecting past gas consumption trends, periodicity and short-term fluctuation characteristics; (2) future covariate sequence, such as weather forecasts and holiday markers, to provide known external information for the future; (3) time resolution category embedding, to control the time granularity of the current prediction task, and to achieve unified modeling of predictions at multiple time resolutions. The above information works together on the denoising network of the diffusion model, so that the model not only has a strong distribution modeling capability, but also can flexibly cope with multiple task settings, improving the generalization ability, uncertainty expression ability and covariate adaptation ability in actual power plant coal consumption prediction scenarios.

[0088] Time resolution refers to the time interval between two adjacent data points in time series data.

[0089] The diffusion model is a powerful probabilistic generative model that learns a mapping from noise to future sequences conditioned on historical data, enabling it to generate high-quality, diverse distributions of future possibilities. It represents a significant paradigm shift in time series forecasting from "guessing the most likely number" to "describing all possible futures."

[0090] Covariates are additional information or context used in time series forecasting to improve model performance and interpretability. Correctly distinguishing and using known and unknown future covariates is key to building a successful and practical forecasting system.

[0091] In the embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative; for example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0092] It should be noted that, in this invention, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element limited by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0093] While the embodiments disclosed in this invention are as described above, the above content is merely for the purpose of facilitating understanding of this invention and is not intended to limit the invention. Any person skilled in the art to which this invention pertains may make any modifications and changes in form and detail of the implementation without departing from the spirit and scope disclosed in this invention; however, the scope of patent protection of this invention shall still be determined by the scope defined in the appended claims.

Claims

1. A method for predicting coal consumption in power plants, characterized in that, include: Step S01: Collect temperature data and power generation data corresponding to the coal consumption of the power plant to obtain the coal consumption dataset of the power plant. Step S02: Divide the power plant coal consumption dataset into multivariate time series, future covariate series, and time target series, and set the resolution parameters; Step S03: Build a covariate conditional diffusion model for predicting coal consumption in power plants over time with multiple resolutions. Step S04: Input the historical time series and future covariate series from the multivariate time series into the conditional channel of the covariate conditional diffusion model, and input the time target series into the target channel of the covariate conditional diffusion model to predict the coal consumption of the power plant.

2. The method according to claim 1, characterized in that, Step S03, the step of building a covariate conditional diffusion model for predicting coal consumption in power plants over time at multiple resolutions, includes: Step S31: Add Gaussian noise to the given original target sequence to construct at least one noisy sample; Step S32: Based on the inverse denoising network, conditional modeling is performed on the noisy samples to recover the original target sequence.

3. The method according to claim 2, characterized in that, The reverse denoising network in step S32 is composed of multiple layers of residual blocks stacked together. Each residual block includes: a diffusion step embedding module, a conditional information fusion module, and a spatiotemporal Transformer module. The diffusion step embedding module encodes the current diffusion step number into a vector through position encoding, and then passes it through a fully connected layer and SiLU activation to obtain the embedded representation of the diffusion step number; The conditional information fusion module maps the resolution category embedding and time step embedding representation to the same dimension as the main network channel, and then extends it to the same length as the input sequence before injecting it into the residual block through addition. The Spatiotemporal Transformer module captures long-range dependencies in sequence time through the time layer and models the interaction structure between multiple variables through the feature layer. The output of the spatiotemporal Transformer module is fed into the residual connection branch and the jump connection branch, and the jump connection results are accumulated layer by layer to obtain the noise prediction.

4. The method according to claim 3, characterized in that, The spatiotemporal Transformer module includes: The time layer is used to capture long-range dependencies of sequences over time; Feature layer, used to model the interaction structure between multiple variables; The outputs of the time layer and the feature layer are co-modeled to obtain the time series modeling result.

5. The method according to claim 1, characterized in that, Step S04, which involves inputting the historical time series and future covariate series from the multivariate time series into the conditional channel of the covariate conditional diffusion model, includes: Step S41: Input the historical time series into the conditional channel, wherein the historical time series reflects past trends and cycles; Step S42: Pass the future covariate sequence into the conditional channel, wherein the future covariate sequence provides known information about the future; Step S43: Input resolution category embedding to control the temporal granularity of the prediction output.

6. The method according to claim 5, characterized in that, Step S43, the step of inputting resolution category embedding, includes: Step S431: Assign a learnable resolution category embedding vector to each temporal resolution category; Step S432: Add the resolution category embedding vector to the time step embedding representation obtained by position encoding of the time step to obtain the fused conditional embedding representation; Step S433: The fused conditional embedding representation is input into each layer of the diffusion denoising network as a global control signal.

7. A power plant coal consumption prediction device, characterized in that, include: The data collection module is used to collect temperature data and power generation data corresponding to the coal consumption of the power plant, and to obtain the coal consumption dataset of the power plant. The data partitioning module is used to partition the power plant coal consumption dataset into a multivariate time series, a future covariate series, and a time target series, and to set resolution parameters. The model building module is used to build a covariate conditional diffusion model for predicting coal consumption in power plants over a time-multi-resolution scale. The prediction module is used to input the historical time series and future covariate series from the multivariate time series into the conditional channel of the covariate conditional diffusion model, and input the time target series into the target channel of the covariate conditional diffusion model to predict the coal consumption of power plants.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

9. An electronic device comprising a processor and a memory, characterized in that, The memory stores a computer program, and the processor executes the computer program to implement the method of any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

Images