A method and related device for multi-time scale timing prediction of a power generation process

Abstract

The present application belongs to the field of intelligent analysis and prediction control of power generation process, and discloses a kind of time sequence prediction method of multi-time scale of power generation process and related device, original data is decomposed into random, periodic and trend component using empirical mode decomposition and wavelet transform, time-frequency and statistical characteristics are extracted, explicit modeling of second-level disturbance, minute-level regulation and hour-level load is realized. Map each component through three layers of latent variable encoder, convert the thermodynamic and mass balance model into a differentiable physical constraint layer, and introduce a physical consistency loss. Use the Transformer submodule to infer multi-scale fusion features, and set up parallel set point prediction, interval prediction and confidence prediction output heads, which are jointly optimized by multi-task loss. Finally, the prediction results are checked for physical consistency, and the illegal values are iteratively corrected in the reverse direction along the gradient. The present application improves the accuracy, reliability and credibility of time series prediction in power generation process, and provides high-precision and high-reliability prediction support for safe and economic operation of thermal power units.

Description

Technical Field

[0001] This invention belongs to the field of intelligent analysis and predictive control technology for power generation processes, and particularly relates to a time-series prediction method and related device for power generation processes across multiple time scales. Background Technology

[0002] With the continuous improvement of the intelligence level of the power system, thermal power units, as the main force of my country's power supply, are directly related to the stable supply and overall energy efficiency of the power system in terms of their operational safety and economy. The operation of thermal power units involves dozens of key parameters, such as main steam temperature, main steam pressure, feedwater temperature, flue gas temperature, load command, fuel flow, and valve opening. There are complex temporal dependencies and spatial coupling relationships among these parameters. Accurate prediction of key operating parameters is an important technical foundation for achieving optimized unit control, safety early warning, and energy efficiency improvement. At present, time-series prediction models for the power generation process are mainly divided into traditional statistical models and deep learning models. Traditional prediction models based on statistical methods are difficult to fully capture the complex coupling relationships between parameters, while prediction models based solely on deep learning face problems such as poor interpretability, weak generalization ability, and difficulty in ensuring physical consistency. Neither of these can fully meet the high safety and high reliability requirements of the power generation industry.

[0003] Current time-series forecasting techniques for power generation still have several shortcomings, specifically: First, they lack explicit modeling of the multi-timescale characteristics of the power generation process. Power generation operating parameters simultaneously encompass characteristics at different time scales, such as hourly load variation cycles, minute-level regulation fluctuations, and second-level disturbance noise. Existing methods generally employ a single-time-scale modeling strategy, failing to effectively extract and model multi-scale features simultaneously. Second, they neglect the integration of prior physical knowledge. Power generation processes follow fundamental physical laws such as thermodynamics and mass balance equations. These physical constraints should ideally be used as prior knowledge for model training to ensure the rationality of predictions. However, most existing methods rely on purely data-driven modeling. First, the methods used are prone to producing predictions that violate basic physical laws and equipment safety boundaries. Second, the predictions lack the ability to quantify uncertainty. In critical power generation safety scenarios, operators not only need the predicted parameter values ​​but also need to clarify the reliability of the prediction results. However, most existing methods only output point prediction values ​​and cannot provide uncertainty information such as prediction intervals and confidence levels to support decision-making. Third, the models lack interpretability. Deep learning models usually have "black box" characteristics, and their internal feature learning and prediction logic is difficult to explain intuitively. In industries with high safety requirements such as power generation, the lack of interpretability will reduce operators' trust in the model's prediction results and limit the engineering application of the model.

[0004] It is evident that existing power generation process time series prediction technologies suffer from problems such as lack of multi-timescale feature modeling, insufficient integration of physical prior knowledge, lack of uncertainty quantification capabilities, and poor model interpretability. These issues fail to meet the application requirements of high-precision, high-reliability, and interpretable time series prediction technologies for the safe and economical operation of thermal power units. Summary of the Invention

[0005] This invention provides a time series prediction method and related apparatus for power generation processes across multiple time scales. This method can effectively solve the problems of existing time series prediction technologies for power generation processes, such as missing multi-time scale feature modeling, insufficient integration of physical prior knowledge, lack of uncertainty quantification capabilities, and poor model interpretability. It can meet the application requirements of high-precision, high-reliability, and interpretable time series prediction technology for the safe and economical operation of thermal power units.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A time-series prediction method for power generation processes across multiple time scales, comprising: Based on the empirical mode decomposition method, the original time series data of multiple operating parameters acquired in real time are decomposed into several intrinsic mode functions. By wavelet transform according to the periodic scale, each intrinsic mode function is merged into random component, periodic component and trend component. At the same time, the statistical features and time-frequency features of each component are extracted. A three-layer latent variable encoder is constructed, corresponding to the trend component, periodic component, and random component respectively. The features of the corresponding components are input into each layer encoder for feature mapping to obtain the trend latent variable, periodic latent variable, and random latent variable. A thermodynamic physical model and a mass balance model of the power generation process are constructed. The physical model is transformed into a differentiable physical constraint layer. The latent variables of each layer and the original observation data are used as inputs to the physical constraint layer to calculate and generate the physical consistency loss term. Based on the physical consistency loss term, the trend latent variable, periodic latent variable and random latent variable are respectively input into the corresponding Transformer sub-module for time series feature inference, and multi-scale fusion inference features are obtained. The multi-scale fusion inference features are input into the point prediction output head, interval prediction output head and confidence prediction output head set in parallel. The three prediction tasks are jointly optimized through a multi-task loss function, and the initial prediction results including point prediction values, prediction intervals and confidence scores are output. The initial prediction results are physically consistent with the preset physical constraint rules. For predictions that fail the check, the results are iteratively corrected in the opposite direction of the physical constraint gradient to obtain the final prediction results.

[0007] Furthermore, the process of acquiring the raw time-series data of the various operating parameters includes: acquiring operating parameters such as main steam temperature, main steam pressure, feedwater temperature, flue gas temperature, load command, fuel flow rate, and valve opening degree in real time from the distributed control system of the thermal power unit; performing quality control on the raw time-series data after acquisition, using the three-standard-deviation criterion to identify outliers exceeding the threshold range, identifying missing data segments with a continuous missing duration exceeding a preset time, and identifying high-frequency oscillation noise data in the time series; correcting outlier data using an interpolation method based on a variational autoencoder, filling missing data using a filling method based on latent variable inference, and using the processed quality-controlled time-series data for empirical mode decomposition.

[0008] Furthermore, the step of merging the intrinsic mode functions into random components, periodic components, and trend components using wavelet transform on a periodic scale specifically involves: The decomposition layer is set to 3 layers, where the period of the high-frequency random component is less than 1 hour, the period of the mid-frequency periodic component is 1 to 24 hours, and the period of the low-frequency trend component is greater than 24 hours; the wavelet transform uses the db4 wavelet basis function; the extracted statistical features include mean, variance, skewness, and kurtosis; the extracted time-frequency features include instantaneous frequency and instantaneous amplitude; the extracted component features are input into the corresponding three-layer latent variable encoder.

[0009] Furthermore, the three-layer latent variable encoders are a trend latent variable encoder, a periodic latent variable encoder, and a random latent variable encoder. The trend latent variable encoder uses a one-dimensional convolutional network to extract trend features, which are then mapped through a fully connected layer to obtain 64-dimensional trend latent variables. The periodic latent variable encoder uses a Fourier transform module to extract periodic features, which are then mapped through a fully connected layer to obtain 32-dimensional periodic latent variables. The random latent variable encoder uses a variational autoencoder framework to encode 64-dimensional random latent variables. The Kolb-Leibler divergence of the latent variables output by each encoder layer is used as a regularization term in model training. The three types of latent variables generated are input into the corresponding Transformer submodule and the physical constraint layer, respectively.

[0010] Furthermore, the thermodynamic physical model includes a boiler water-cooled wall heat transfer model, a superheater heat transfer model, and a steam state equation; physical constraint terms are added to the loss function using the Lagrange multiplier method, and the weights of the physical constraints are dynamically adjusted through learnable parameters; the calculated physical consistency loss term is added to the multi-task loss function and participates in the model training and optimization process together with the prediction task loss.

[0011] Furthermore, the Transformer submodules are a Trend Transformer submodule, a Periodic Transformer submodule, and a Random Transformer submodule. The Trend Transformer submodule employs a 12-layer full attention structure, with 8 attention heads per layer and a latent variable dimension of 512. The Periodic Transformer submodule employs a 6-layer periodic-aware sparse attention structure, with 8 attention heads per layer and a latent variable dimension of 256. The Random Transformer submodule employs a 6-layer standard self-attention structure, with 8 attention heads per layer and a latent variable dimension of 512. The three submodules interact with each other through a cross-layer attention mechanism. The weights of the cross-layer attention are determined by a learnable parameter matrix, ultimately fusing to generate multi-scale fusion inference features.

[0012] Furthermore, the point prediction output head uses a two-layer fully connected network to output the predicted value for the next 24 hours, the interval prediction output head uses a quantile regression method to output prediction intervals for five quantiles: 10%, 25%, 50%, 75%, and 90%, and the confidence output head uses a Bayesian linear regression layer to output the prediction variance estimate. The three prediction tasks are jointly optimized through a multi-task loss function with dynamic weights. The physical consistency verification includes temperature range check, pressure range check, thermodynamic relationship check, and mass balance check. Predicted values ​​that fail the verification are iteratively corrected with a preset correction step size, and the final prediction result is obtained and output after correction.

[0013] A time-series prediction system for a power generation process with multiple time scales, comprising: The decomposition module is used to decompose the raw time-series data of real-time acquisition of multiple operating parameters into several intrinsic mode functions based on the empirical mode decomposition method. The intrinsic mode functions are then merged into random components, periodic components and trend components by wavelet transform according to the periodic scale, while extracting the statistical features and time-frequency features of each component. The mapping module is used to construct a three-layer latent variable encoder that corresponds to the trend component, periodic component, and random component, respectively. The features of the corresponding components are input into each layer encoder for feature mapping to obtain the trend latent variable, periodic latent variable, and random latent variable. The calculation module is used to construct the thermodynamic physical model and mass balance model of the power generation process, transform the physical model into a differentiable physical constraint layer, and use the latent variables of each layer and the original observation data as inputs to the physical constraint layer to calculate and generate the physical consistency loss term. The inference module is used to perform time-series feature inference by inputting the trend latent variable, periodic latent variable and random latent variable into the corresponding Transformer sub-module based on the physical consistency loss term, and obtain multi-scale fusion inference features. The prediction module is used to input multi-scale fused inference features into the point prediction output head, interval prediction output head and confidence prediction output head set in parallel. The three prediction tasks are jointly optimized through a multi-task loss function, and the output includes the point prediction value, prediction interval and confidence score. The correction module is used to perform physical consistency verification on the initial prediction results based on preset physical constraint rules. For prediction values ​​that fail the verification, iterative correction is performed along the reverse direction of the physical constraint gradient to obtain the final prediction result.

[0014] A time-series prediction device for a power generation process across multiple time scales, comprising: Memory, used to store computer programs; A processor is used to implement the above-described time-scale prediction method for the power generation process when executing the computer program.

[0015] A computer-readable storage medium storing a computer program, which, when executed by a processor, is used to implement the aforementioned multi-timescale time-scale prediction method for the power generation process.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a multi-timescale time-series prediction method for power generation processes. It utilizes empirical mode decomposition and wavelet transform to decompose the original time-series data into random, periodic, and trend components, extracting their time-frequency and statistical features. This enables explicit modeling of multi-timescale features of the power generation process, such as second-level disturbances, minute-level adjustments, and hourly loads. Subsequently, a three-layer latent variable encoder is constructed to map each component to its corresponding latent variable. The thermodynamic physical model and mass balance model are transformed into a differentiable physical constraint layer, introducing a physical consistency loss term to ensure the model automatically follows fundamental physical laws and safety boundaries during the learning process. Based on this, a Transformer submodule is used to perform time-series inference on each latent variable to obtain multi-scale fused features. Parallel point prediction, interval prediction, and confidence prediction output heads are used, and multi-task loss is jointly optimized to output point prediction values, prediction intervals, and confidence scores. Finally, the initial prediction results are physically consistent, and violations are iteratively corrected along the physical constraint gradient. This method effectively overcomes the shortcomings of existing technologies, such as lack of multi-scale feature modeling, insufficient integration of physical priors, lack of uncertainty quantification, and poor interpretability. It significantly improves the accuracy, reliability, and credibility of power generation process time series prediction, providing high-precision, high-reliability, and interpretable prediction support for the safe and economical operation of thermal power units. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a hierarchical latent variable time series prediction architecture provided in an embodiment of the present invention; Figure 2 This is a flowchart of physical constraint injection and verification provided in an embodiment of the present invention; Figure 3 The following is a comparison chart of the main steam temperature prediction performance of a 600MW unit provided in the embodiments of the present invention; wherein, (a) is a comparison of MAE index; (b) is a comparison of RMSE index; (c) is a comparison of 90% prediction interval coverage; and (d) is a comparison of physical consistency score. Figure 4 A flowchart illustrating a multi-timescale time-series prediction method for power generation processes provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a multi-timescale time-series prediction system for power generation processes provided in an embodiment of the present invention. Detailed Implementation

[0018] To further understand the content of this invention, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention.

[0019] As described in the background section, current time-series prediction techniques for power generation processes have the following shortcomings: First, existing methods lack explicit modeling of the multi-timescale characteristics of the power generation process. The operating parameters of the power generation process include load variation cycles measured in hours, regulation fluctuations measured in minutes, and disturbance noise measured in seconds. Existing methods often employ a single-timescale modeling strategy, making it difficult to capture these features at different scales simultaneously.

[0020] Second, existing methods neglect the incorporation of prior physical knowledge. The power generation process follows physical laws such as thermodynamics and mass balance equations. These physical constraints should be considered prior knowledge for model learning, but most existing methods adopt a purely data-driven approach, which may lead to prediction results that violate physical laws.

[0021] Third, existing methods lack uncertainty quantification in their predictions. In power generation safety-critical scenarios, operators need not only the predicted values ​​but also their reliability. Most existing methods only output point predictions, failing to provide prediction intervals and confidence levels.

[0022] Fourth, existing methods suffer from insufficient model interpretability. Deep learning models are often considered "black box" models, with their internal working mechanisms difficult to explain. In industries with high safety requirements, such as power generation, model interpretability is crucial for enhancing operators' trust in the models.

[0023] To address the aforementioned issues, this embodiment provides a time-series prediction method for power generation processes across multiple time scales. Specifically, this method achieves the following objectives: Objective 1: Explicit modeling across multiple time scales through a data decomposition layer; Objective 2: End-to-end injection of physical priors through a physical constraint layer; Objective 3: Uncertainty quantification through a multi-task output layer; Objective 4: Dynamic parameter configuration through an adaptive operating condition mechanism; Objective 5: Continuous optimization through an incremental learning module; Objective 6: Efficient parallel training through a distributed training module; Objective 7: Automatic result correction through a physical consistency verification module.

[0024] For example, such as Figure 4 As shown, this embodiment provides a time-series prediction method for power generation processes across multiple time scales, including: Based on the empirical mode decomposition method, the original time series data of multiple operating parameters acquired in real time are decomposed into several intrinsic mode functions. By wavelet transform according to the periodic scale, each intrinsic mode function is merged into random component, periodic component and trend component. At the same time, the statistical features and time-frequency features of each component are extracted. A three-layer latent variable encoder is constructed, corresponding to the trend component, periodic component, and random component respectively. The features of the corresponding components are input into each layer encoder for feature mapping to obtain the trend latent variable, periodic latent variable, and random latent variable. A thermodynamic physical model and a mass balance model of the power generation process are constructed. The physical model is transformed into a differentiable physical constraint layer. The latent variables of each layer and the original observation data are used as inputs to the physical constraint layer to calculate and generate the physical consistency loss term. Based on the physical consistency loss term, the trend latent variable, periodic latent variable and random latent variable are respectively input into the corresponding Transformer sub-module for time series feature inference, and multi-scale fusion inference features are obtained. The multi-scale fusion inference features are input into the point prediction output head, interval prediction output head and confidence prediction output head set in parallel. The three prediction tasks are jointly optimized through a multi-task loss function, and the initial prediction results including point prediction values, prediction intervals and confidence scores are output. The initial prediction results are physically consistent with the preset physical constraint rules. For predictions that fail the check, the results are iteratively corrected in the opposite direction of the physical constraint gradient to obtain the final prediction results.

[0025] The time-series prediction method for the power generation process at multiple time scales provided in this embodiment will be further explained below with reference to the accompanying drawings: For example, this embodiment provides a time-series prediction method for power generation processes across multiple time scales. This embodiment employs a Hierarchical Latent Variable Architecture (HLVA) to decompose the time-series data of the power generation process into three levels: a trend layer, a periodic layer, and a stochastic layer. These levels are encoded into corresponding latent variable spaces, and then time-series inference is performed using a hierarchical Transformer. The specific steps are as follows: Step 1: Time Series Data Acquisition and Quality Control Specific operations: Real-time acquisition of operating parameter data from the thermal power unit's DCS system, including key parameters such as main steam temperature T_ms (°C), main steam pressure P_ms (MPa), feedwater temperature T_fw (°C), flue gas temperature T_flue (°C), load command L (MW), fuel flow rate F_fuel (t / h), and valve opening V (%). The acquisition frequency is set to 1 second / time to ensure the temporal integrity of the data. Quality control is performed on the acquired data, automatically identifying outliers (data points exceeding 3σ), missing values ​​(data segments missing for more than 10 seconds consecutively), and noise data (high-frequency oscillation components). For identified outliers, an interpolation method based on a variational autoencoder is used for correction; for missing data, an imputation method based on latent variable inference is used for imputation.

[0026] Key parameters: Acquisition frequency = 1 second / time; Outlier detection threshold = 3σ; Continuous missing alarm threshold = 10 seconds; Variational autoencoder interpolation latent dimension = 16 dimensions.

[0027] Judgment criteria: Pass standard - data integrity rate ≥ 99.5% (percentage of valid data points within 24 hours); Failure standard - data source check is triggered when data integrity rate < 99.5%, and system alarm is triggered when missing rate > 2%.

[0028] Step 2: Multi-scale time-series signal decomposition: Specific procedures: The Empirical Mode Decomposition (EMD) method is used to decompose the original time-series signal into several Intrinsic Mode Functions (IMFs). Then, wavelet transform is used to further merge high-frequency IMFs into random components and low-frequency IMFs into trend components. The decomposition layer is set to three levels: high-frequency layer (random components, period less than 1 hour), mid-frequency layer (periodic components, period 1 to 24 hours), and low-frequency layer (trend components, period greater than 24 hours). Statistical characteristics (mean, variance, skewness, kurtosis) and time-frequency characteristics (instantaneous frequency, instantaneous amplitude) of each component are extracted to prepare for subsequent latent variable encoding.

[0029] Key parameters: EMD decomposition layer number = 3 layers; high frequency layer period threshold = <1 hour; mid frequency layer period range = 1 to 24 hours; low frequency layer period threshold => 24 hours; wavelet basis function = db4; decomposition iteration convergence threshold = 0.05.

[0030] Judgment criteria: Pass criterion - reconstruction error of each component after decomposition <1% (relative error between the original signal and the component superposition); Failure criterion - if the decomposition does not converge or the reconstruction error is ≥1%, change the decomposition method (use variational mode decomposition instead).

[0031] Step 3: Hierarchical latent variable coding: Specific operations: A three-layer latent variable encoder is constructed, corresponding to the trend, periodic, and random time series levels, respectively. The trend latent variable encoder uses a one-dimensional convolutional network (kernel size 3, stride 1, number of channels 64) to extract trend features, and then maps them to a 64-dimensional trend latent variable space through a fully connected layer. The periodic latent variable encoder uses a Fourier transform module to extract periodic features, and then maps them to a 32-dimensional periodic latent variable space through a fully connected layer. The random latent variable encoder uses a variational autoencoder framework, including a mean predictor and a variance predictor, to encode random features into 64-dimensional random latent variables (following a standard normal distribution). The KL divergence of the latent variables output by each encoder layer is used as a regularization term.

[0032] Key parameters: Trend latent variable dimension = 64; Periodic latent variable dimension = 32; Random latent variable dimension = 64; Convolution kernel size = 3; Convolution stride = 1; Number of channels = 64; KL divergence weight = 0.01; Variational inference sampling times = 1 (training) / 100 (inference).

[0033] Judgment criteria: Pass criterion - convergence of latent variable reconstruction test error at each layer (error decrease of <0.1% over 5 consecutive epochs); Failure criterion - when insufficient latent variable dimensions lead to reconstruction error >5%, dimension expansion is triggered, adding 16 dimensions.

[0034] Step 4: Physical constraint modeling and coding: Specific steps: Construct a thermodynamic physical model of the power generation process, including: boiler water-cooled wall heat transfer model Q_wall = κ_wall · A_wall · (T_flue) T_water), superheater heat transfer model Q_sh=κ_sh·A_sh·(T_gas) The physical models include the steam equation of state f(P_ms,T_ms,v_s)=0 and the mass balance equation. Differentiable approximations of these physical models are embedded into a neural network to construct a physical constraint layer. The input to the physical constraint layer is the latent variables and the original observation data, and the output is the physical consistency loss term.

[0035] Key parameters: physical constraint weight λ_phys=0.1; initial value of Lagrange multiplier=1.0; heat transfer coefficients κ_wall and κ_sh are learned through data-driven learning; iterative solution step size of state equation=0.01; physical constraint loss calculation batch=each training batch.

[0036] Judgment criteria: Pass criterion - physical constraint loss term converges to <0.01 (after normalization); Failure criterion - if physical constraint loss is consistently >0.1, check whether the physical model parameters are reasonable and adjust the range of λ_phys to [0.05, 0.5].

[0037] Step 5: Hierarchical Transformer Inference: Specific operation: The three layers of latent variables are input into their respective Transformer sub-modules for inference. The Trend Transformer employs a 12-layer full attention structure, with each layer containing 8 attention heads and a latent variable dimension of 512, used to capture long-distance dependencies of trend components; the Periodic Transformer employs a 6-layer period-aware sparse attention structure, with each layer containing 8 attention heads and a latent variable dimension of 256, using period-aware relative position encoding; the Stochastic Transformer employs a 6-layer standard self-attention structure, with each layer containing 8 attention heads and a latent variable dimension of 512, used to model random perturbations. The three sub-modules interact with each other through a cross-layer attention mechanism, and the weights of the cross-layer attention are determined by a learnable parameter matrix.

[0038] Key parameters: Trend Transformer layers = 12; Periodic Transformer layers = 6; Random Transformer layers = 6; Number of attention heads = 8; Trend Transformer latent variable dimension = 512; Periodic Transformer latent variable dimension = 256; Random Transformer latent variable dimension = 512; Feedforward network scaling factor = 4; Dropout rate = 0.1; Position encoding = learnable periodic-aware relative position encoding.

[0039] Judgment criteria: Pass criterion – stop training when the MAE predicted on the validation set shows no decreasing trend for 10 consecutive epochs; Failure criterion – reduce the learning rate to 1 / 10 of the current value and restart training when the training loss does not converge (gradient explosion).

[0040] Step 6: Multi-task prediction output: Specific operation: Three parallel output heads are designed to correspond to the point prediction, interval prediction, and confidence prediction tasks, respectively. The point prediction output head uses a two-layer fully connected network (hidden layer dimension 256, output dimension = prediction step size) to directly output the predicted value for the next 24 hours; the interval prediction output head uses a quantile regression method to output the predicted values ​​for five quantiles: 10%, 25%, 50%, 75%, and 90%; the confidence output head uses a Bayesian linear regression layer to output the variance estimate of the prediction. The three tasks share the underlying feature extractor and are jointly optimized using the multi-task loss function L_total = L_point + α·L_interval + β·L_conf, where α and β are dynamic weights.

[0041] Key parameters: Hidden layer dimension = 256; Prediction step size = 24 hours; Quantile regression quantiles = [0.1, 0.25, 0.5, 0.75, 0.9]; Bayesian inference prior variance = 1.0; Dynamic weight update cycle = every 10 batches; Initial α = 1.0; Initial β = 0.5.

[0042] Judgment criteria: Pass criterion - the loss weights of the three tasks converge and stabilize (variance < 0.01); Failure criterion - when the loss weight of a certain task approaches 0 (is ignored), the minimum weight of that task is forcibly set to 0.3.

[0043] Step 7: Physical Consistency Verification and Correction: Specific procedures: Perform physical consistency verification on the predicted values ​​output by the model. Verification includes: temperature range check (T_ms∈[500°C,600°C]), pressure range check (P_ms∈[16MPa,25MPa]), thermodynamic relationship check (ΔT_ms should match the heat load change), and mass balance check (feedwater flow rate should match steam flow rate). For predicted values ​​that fail verification, a correction algorithm based on the physical model is used: calculate the gradient direction of the physical constraints, iterate along the opposite direction of the gradient with a correction step size until the predicted value meets the physical constraints or the number of iterations reaches the upper limit.

[0044] Key parameters: Main steam temperature effective range = [500°C, 600°C]; Main steam pressure effective range = [16MPa, 25MPa]; Correction step size = 0.01; Maximum number of iterations = 50; Physical consistency score threshold = 0.8.

[0045] Judgment criteria: Pass criteria - Physical consistency score ≥ 0.8 and the corrected prediction value falls within the valid range; Failure criteria - If the constraint is still not met after 50 corrected iterations, the prediction is marked as low confidence, the original prediction value is output and a warning label is attached.

[0046] Step 8: Distributed Model Training Specific operations: A parameter server architecture is used for distributed model training. The master node maintains a complete copy of the model parameters, while each worker node is responsible for calculating local gradients. The training data is divided into multiple time windows in chronological order, with each window containing 168 hours of historical data and 24 hours of prediction targets. Each worker node uses different data windows to calculate gradients in parallel, and the master node performs gradient aggregation and parameter updates every 10 batches. Gradient compression techniques (Top-K sparsity, compression rate of 90%) and gradient delay compensation techniques (error feedback mechanism) are implemented to reduce communication overhead.

[0047] Key parameters: Time window size = 168 hours of history + 24 hours of prediction; Gradient aggregation cycle = 10 batches; Top-K sparsity compression rate = 90%; Initial learning rate = 1e-4; Final learning rate = 1e-6 (cosine annealing); Total training epochs = 100 epochs; Batch size = 32; Number of worker nodes = 4 nodes (expandable to 16 nodes).

[0048] Judgment criteria: Pass criteria - training is terminated when the validation set MAE < 0.35°C and the physical consistency score > 0.9; Failure criteria - gradient clipping is triggered (threshold = 1.0) and the learning rate is reduced when the training loss diverges (single batch gradient norm > 100).

[0049] Step 9: Incremental Learning and Model Update Specific operations: After model deployment, continuously collect new data and perform incremental learning updates. The incremental learning cycle is set to 24 hours, meaning the model is updated daily using the latest 24-hour data. Incremental learning employs an experience replay strategy, maintaining an experience pool of 10,000 samples. Each update randomly samples 50% new data and 50% historical data from this pool. Concept drift is detected by comparing the mean and variance of the prediction errors over the past 7 days with the historical baseline. When a significant difference is detected (KS test p-value < 0.05), full retraining is triggered. During incremental updates, the core model parameters remain unchanged; only the subset of parameters relevant to the new data is updated.

[0050] Key parameters: Incremental learning period = 24 hours; Experience pool capacity = 10,000 records; New and old data sampling ratio = 50%:50%; KS test significance level = 0.05; Concept drift detection window = 7 days; Parameter subset update ratio = 30%.

[0051] Judgment criteria: Pass criterion - No significant increase in verification error after incremental update (error increase <5%); Failure criterion - Triggering a full retraining process when concept drift is detected (p value <0.05), shortening the update cycle to 6 hours.

[0052] Step 10: Configuration of adaptive operating conditions parameters: Specific operation: The model parameters are dynamically adjusted based on the identified operating conditions. The operating condition identifier comprehensively considers indicators such as load change rate (|dL / dt|, unit: % / min), temperature change rate (|dT_ms / dt|, unit: °C / min), and pressure change rate (|dP_ms / dt|, unit: MPa / min), and outputs operating condition category labels. The identified operating condition types include 11 basic operating conditions such as: start-up, shutdown, stable operation at rated load (load change rate <1% / min), variable load operation (±5% / min), deep peak shaving (40% load), unstable combustion, feedwater failure, load disturbance, fuel fluctuation, thermal deviation, and normal stability, as well as 2 CFB-specific operating conditions: abnormal fluidization state and abnormal return material circulation. Each operating condition corresponds to an independent subset of parameter configurations.

[0053] Key parameters: Load change rate threshold = 1% / min (steady-state / dynamic boundary); Temperature change rate threshold = 0.5°C / min; Pressure change rate threshold = 0.1MPa / min; Hidden layer dimension of the working condition identification model = 128; Proportion of independent parameters for each working condition = 20% to 40%.

[0054] Judgment criteria: Pass standard - working condition recognition accuracy ≥ 85% (verified based on historical labeled data); Failure standard - recognition accuracy < 70% will revert to default parameter configuration and trigger an alarm.

[0055] Step 11: Prediction Result Output and Alarms: Specific operations: Output the final prediction results in a standardized format, including: prediction timestamp, prediction parameter list, predicted values ​​for each parameter, prediction interval (10% to 90th percentile), confidence score, physical consistency score, and operating condition category. Set three alarm thresholds: Level 1 alarm (prediction deviation > 5°C or confidence < 0.6, alarm level "Severe"), Level 2 alarm (prediction deviation 3 to 5°C or confidence 0.6 to 0.8, alarm level "Warning"), and Level 3 alarm (prediction deviation 1 to 3°C or confidence 0.8 to 0.9, alarm level "Alert"). Alarm information is pushed to the power plant monitoring center through the system interface, and alarm history is recorded for subsequent analysis.

[0056] Key parameters: Level 1 alarm deviation threshold = 5°C; Level 2 alarm deviation threshold range = 3 to 5°C; Level 3 alarm deviation threshold range = 1 to 3°C; Level 1 confidence threshold = 0.6; Level 2 confidence threshold range = 0.6 to 0.8; Level 3 confidence threshold range = 0.8 to 0.9; Alarm push delay < 1 second.

[0057] Judgment criteria: Pass criteria - alarm trigger response time < 1 second, alarm information integrity rate = 100%; Fail criteria - alarm delay > 5 seconds, downgrade to asynchronous push mode, and record alarm backlog.

[0058] Step 12: Model Version Management and Canary Release: Specific operations: After the new model is trained, it is managed through the version management module. The new model is first deployed and validated on a small scale on a gray-scale node (10% traffic). After successful validation, traffic is gradually switched to the new model (20%, 50%, 100%). The model files, configuration parameters, and performance metrics for each version are recorded, supporting rapid rollback. Model health checks include: prediction latency <50ms, stable memory usage, and a physical consistency score consistently >0.85.

[0059] Key parameters: Gray release traffic ratio = [10%, 20%, 50%, 100%]; Health check cycle = every 5 minutes; Predicted latency alarm threshold = 50ms; Memory usage growth threshold = 20%; Physical consistency score alarm threshold = 0.85; Number of versions retained = 10.

[0060] Judgment criteria: Pass standard - all grayscale node health checks pass (latency, memory, and physical score all meet the threshold); Failure standard - if any health check indicator fails to meet the standard for 3 consecutive times, the system will automatically roll back to the previous version.

[0061] For example, in order to implement the time series prediction method provided in the above embodiments, a hardware structure for a multi-time-scale time series prediction system for power generation is also provided. This system consists of 13 functional modules, and the modules interact with each other through standardized interfaces to form a complete time series prediction processing pipeline, as detailed below: Module 1: Time Series Data Acquisition and Preprocessing Module Function Description: Responsible for real-time acquisition of operating parameter data from the distributed control system (DCS) of thermal power units, and performing data cleaning, missing value filling and standardization processing.

[0062] Input: Real-time data stream from the DCS system, including parameters such as main steam temperature T_ms (unit: °C), main steam pressure P_ms (unit: MPa), feedwater temperature T_fw (unit: °C), flue gas temperature T_flue (unit: °C), load command L (unit: MW), fuel flow rate F_fuel (unit: t / h), and valve opening V (unit: %).

[0063] Output: Standardized time series data matrix (each row corresponds to a time step, and each column corresponds to a parameter).

[0064] Interface relationships: It connects to the multi-scale time series decomposition module through a data pipeline, and its output serves as the input to the decomposition module; it shares a data index with the physical consistency verification module.

[0065] Module 2: Multi-scale temporal decomposition module: Functional Description: This method combines empirical mode decomposition and wavelet transform to decompose the original time series signal into three levels: trend component, periodic component, and random component.

[0066] Input: Standardized raw time series data.

[0067] Output: Three independent component sequences: trend component (period greater than 24 hours), periodic component (period from 1 to 24 hours), and random component (period less than 1 hour).

[0068] Interface relationship: The input comes from the data acquisition and preprocessing module; the output is connected to three sub-encoders of the hierarchical latent variable coding module.

[0069] Module 3: Hierarchical Latent Variable Encoder Module: Function Description: The decomposed time series data is encoded into three independent latent variable spaces: trend, periodic, and random. Mutual information is maximized to ensure that the information of each layer is complementary.

[0070] Input: Trend component sequence, periodic component sequence, random component sequence.

[0071] Output: Trend latent variable matrix (64-dimensional), Periodic latent variable matrix (32-dimensional), Random latent variable matrix (64-dimensional).

[0072] Interface relationships: Input comes from the multi-scale decomposition module; output is connected to three sub-modules of the hierarchical Transformer inference module; latent variables are simultaneously input into the physical constraint injection module.

[0073] Module 4: Physical Constraint Injection Module Function Description: The thermodynamic laws and mass balance equations of the power generation process are encoded into learnable constraints and integrated into the model training through the Lagrange multiplier method.

[0074] Inputs: latent variable matrix (D-dimensional), raw observation data, and physical constraint configuration parameters.

[0075] Output: Physical consistency loss term (a scalar of gradients that can be backpropagated).

[0076] Interface relationship: Receives the output of the latent variable encoder as input; the loss term is connected to the overall loss function; the physical model parameters can be learned and updated.

[0077] Module 5: Hierarchical Transformer Inference Module Functional description: It uses three parallel Transformer sub-modules to process trend, periodic and random latent variables respectively, and fuses information from each layer through a cross-layer attention mechanism.

[0078] Input: Trend latent variable matrix, periodic latent variable matrix, random latent variable matrix.

[0079] Output: The fused inference feature vector (1536 dimensions, i.e., 512+256+512+256 dimensional fused output).

[0080] Interface relationships: Input comes from the latent variable encoder; output is connected to the multi-task prediction output module; cross-layer attention weights are determined by learnable parameters.

[0081] Module 6: Multi-task prediction output module: Function Description: Simultaneously outputs the results of three tasks: point prediction value, prediction interval, and confidence score, using a dynamic weighted multi-task learning framework.

[0082] Input: Layered Transformer fused feature vector.

[0083] Output: Point prediction (predicted sequence for the next 24 hours), quantile prediction (five quantiles: 10% / 25% / 50% / 75% / 90%), confidence level (prediction variance).

[0084] Interface relationship: Input comes from the inference module; output connects to the physical consistency verification module and the visual alarm module.

[0085] Module 7: Physical Consistency Verification and Correction Module Function Description: Performs physical rationality checks on the prediction results to ensure that the predicted values ​​meet the basic laws of thermodynamics and the safety boundary conditions of the equipment, and automatically corrects the predicted values ​​that violate physical constraints.

[0086] Input: Multi-task prediction output (point prediction value, quantile, confidence score), physical constraint model.

[0087] Output: Corrected prediction results, physical consistency score (between 0 and 1).

[0088] Interface relationship: Receives the output of the multi-task output module; sends the output to the visualization alarm module; and sends the anomaly marker data back to the data acquisition module.

[0089] Module 8: Adaptive Adjustment Module for Operating Conditions Function Description: Identifies the current operating condition type of the unit and dynamically adjusts the model parameter configuration based on the operating condition characteristics.

[0090] Input: Raw time series data, load change rate (unit: % / min), temperature change rate (unit: °C / min), pressure change rate (unit: MPa / min), and other operating condition indicators.

[0091] Output: Working condition category label (one of 13 categories), model parameter adaptation configuration.

[0092] Interface relationships: Receives real-time data from the data acquisition module; outputs adjustment signals to each model parameter configuration node; supports online learning and updating of the working condition recognition model.

[0093] Module 9: Distributed Training Scheduling Module Function Description: Coordinates model training tasks across multiple computing nodes, enabling gradient aggregation, parameter synchronization, and training monitoring.

[0094] Inputs: training data shards, model parameter gradients, and worker node status.

[0095] Output: Aggregated parameter updates and global model parameter copies.

[0096] Interface relationships: bidirectional connection with all model parameter nodes; supports dynamic addition and removal of computation nodes; outputs the final model to the inference module.

[0097] Module 10: Incremental Learning Update Module Function Description: After model deployment, it continuously learns from new data and achieves online iterative optimization of the model through experience replay, knowledge distillation, and concept drift detection.

[0098] Input: New time series data, current model parameters, historical experience pool.

[0099] Output: Updated model parameters and concept drift detection results.

[0100] Interface Relationships: Receives new data; works in collaboration with the distributed training module; triggers a full retraining signal to the scheduling module.

[0101] Module 11: Prediction Result Visualization and Alarm Module Function Description: Presents prediction results to operators in a multi-view format and automatically triggers multi-level alarms based on alarm thresholds.

[0102] Inputs: Corrected prediction results, physical consistency score, and operating condition category.

[0103] Output: Visual interface updates, alarm information push (audio-visual / SMS / email).

[0104] Interface relationships: Receives outputs from the physical verification module and the operating condition identification module; interfaces with the power plant monitoring system; stores alarm history in the database.

[0105] Module 12: Time Series Database Management Module Function Description: Manages and maintains historical time-series data, providing efficient data query and playback functions.

[0106] Inputs: Real-time collected data, labeled data, and imported external data.

[0107] Output: Structured time-series data storage and data slice query results.

[0108] Interface relationships: bidirectional synchronization with the data acquisition module; provides data playback for the training module; supports data export to analysis tools.

[0109] Module 13: Model Version Management and Deployment Module Function Description: Manages multiple versions of a model, supporting model rollback, canary release, and online switching.

[0110] Input: Newly trained model file, deployment strategy configuration, health check results.

[0111] Outputs: Deployment status report, model switching instructions, and monitoring metrics.

[0112] Interface relationships: Receives the newly trained model; connects to the inference module; supports synchronous updates with multiple inference nodes.

[0113] As can be seen, through the collaborative work of the above 13 modules, this hardware system constructs a complete closed-loop system from data acquisition, multi-scale decomposition, hierarchical encoding, physical constraint injection, hierarchical inference, multi-task output, to physical verification, operating condition adaptation, distributed training, incremental learning, visual alarm, data management and model version management, realizing high-precision, high-reliability, multi-time-scale time series prediction of the power generation process.

[0114] In summary, this embodiment provides a time-series prediction method for power generation processes across multiple time scales, with the following specific technical advantages: First, hierarchical latent variable architecture design: This method proposes for the first time a hierarchical latent variable architecture suitable for time-series prediction of power generation processes. It decomposes time-series data into three levels: trend, periodic, and stochastic, with each level corresponding to an independent latent variable space and a dedicated Transformer inference module. Compared to traditional single-level latent variable models, the hierarchical architecture can more accurately capture the multi-timescale characteristics of the power generation process. Trend latent variables specifically encode the long-term aging trends of equipment and the general trend of load changes; periodic latent variables specifically encode periodic features such as diurnal cycles and load cycles; and stochastic latent variables specifically encode random disturbances such as load fluctuations and fuel quality changes. This clearly defined design avoids mutual interference between different features, significantly improving prediction accuracy. Experiments show that compared to the single-level architecture, the hierarchical latent variable architecture reduces the predicted main steam temperature (MAE) from 0.42°C to 0.31°C, a reduction of 26%.

[0115] It should be noted that, as Figure 1 As shown, this embodiment uses the LST-SAM module to achieve time series decomposition. The trend latent variable has 64 dimensions (12 layers of Transformer encoding), the periodic latent variable has 32 dimensions (6 layers of Transformer encoding), and the random latent variable has 64 dimensions (6 layers of Transformer encoding). After parallel encoding of the three components, an attention-weighted fusion layer (256×256 weight matrix) outputs a 256-dimensional fused feature. The data source is a 600MW supercritical thermal power unit DCS system, with a sampling frequency of 1 second / time, and five-dimensional inputs: main steam temperature / pressure / feed water temperature / flue gas temperature / load command. Figure 1 The left side shows the three components of the EMD decomposition node output (distinguished by blue / green / orange lines), the middle shows three parallel Transformer encoding branches (labeled with layer number and dimension), and the right side shows the fusion layer and prediction output header, with gray arrows indicating the data flow. Compared to other traditional methods, this method differs in that: the three components are encoded independently instead of by a single encoder, the fusion weights are learnable rather than simply concatenated, and the fusion layer includes a physical constraint verification node (L_phys weight 0.1)). Figure 1 In the diagram, the input on the left is the original time-series signal, which is divided into three branches after passing through a data decomposition layer: a trend branch, a periodic branch, and a stochastic branch. The trend branch is processed by a trend latent variable encoder and a trend Transformer; the periodic branch is processed by a periodic latent variable encoder and a periodic Transformer; and the stochastic branch is processed by a random latent variable encoder and a stochastic Transformer. The outputs of the three branches are fused through a cross-layer attention mechanism, and finally, the predicted point value, prediction interval, and confidence level are simultaneously output through a multi-task output head. Figure 1 The dashed line represents the physical constraint injection path, and the thermodynamic model output is connected to the loss function.

[0116] Second, the physical constraint end-to-end injection mechanism: This method proposes to embed prior physical knowledge, such as the thermodynamic laws and mass balance equations of the power generation process, into the model training process end-to-end in a differentiable manner. Traditional Physical Information Neural Networks (PINNs) use physical constraints as an additional term in the loss function, but this method struggles to handle complex physical constraint relationships. The innovation of this method lies in constructing a differentiable physical model layer, expressing physical constraints as learnable parameters of the neural network, achieving a seamless integration of physical constraints and data-driven learning. The physical constraint injection includes three levels: energy conservation constraints (balance between boiler heat exchange and steam heat absorption), mass conservation constraints (balance between feedwater flow and steam flow), and thermodynamic state constraints (steam state equation constraints). The introduction of physical constraints not only improves the physical consistency of predictions (physical consistency score increases from 0.62 to 0.97) but also significantly enhances the model's generalization ability in scenarios with few samples.

[0117] It should be noted that, as Figure 2 As shown, in this method, thermodynamic constraints are injected through a custom loss function term, including three sub-terms: energy balance constraint L_energy, temperature-pressure monotonicity constraint L_mono, and mass conservation constraint L_mass, with a weight λ_phys=0.1. The physical verification data comes from a historical database of steady-state operating conditions of 600MW units (≥10,000 records), and the verified physical quantities include four categories: temperature, pressure, enthalpy, and flow rate. Figure 2 The left side is the prediction output node (yellow), the middle is the physical constraint calculation node (the three rectangles represent L_energy / L_mono / L_mass respectively), and the right side is the closed-loop correction node. The correction amount feedback path is marked with a red dashed line, and the correction threshold is marked as ±3σ. Figure 2 It demonstrates end-to-end closed-loop correction rather than open-loop constraints, and includes cross-physical quantity coupling verification (L_energy involves both temperature and pressure). Figure 2 The upper part of the algorithm is the physical constraint injection process: thermodynamic prior knowledge (boiler heat transfer model, superheater model, steam state equation) is converted into a learnable physical constraint layer for the neural network after being processed by differentiable approximation. This physical constraint layer, along with latent variables and observation data, is input into the overall loss function, and joint optimization is achieved through the Lagrange multiplier method. The lower part is the physical consistency verification process: the predicted output undergoes temperature range checks, pressure range checks, thermodynamic relationship checks, and mass balance checks. For predictions that fail the checks, correction iterations are performed along the gradient in the opposite direction until the physical constraints are met or the maximum number of iterations is reached.

[0118] Third, the multi-task prediction output framework: This method designs a multi-task learning framework that simultaneously outputs point predictions, interval predictions, and confidence predictions. Traditional time-series forecasting models only output point predictions, which cannot meet the needs of critical power generation safety scenarios for uncertainty information. The innovation of this method lies in designing three parallel output heads that share the underlying feature representation, and achieving an effective balance among the three tasks through a dynamic weight adjustment strategy. Interval prediction uses a quantile regression method to output predicted values ​​at five quantiles: 10%, 25%, 50%, 75%, and 90%, providing operators with a reference for the possible distribution range of predicted values. Confidence prediction uses a Bayesian inference method to estimate the variance of each prediction point, reflecting the model's confidence in the prediction. This multi-task design enables the model to more comprehensively support operational decisions in the power generation process, increasing the prediction interval coverage (90%) from 72.1% to 91.2%.

[0119] Fourth, configuration of model parameters for adaptive operating conditions: This method proposes an adaptive mechanism that dynamically adjusts model parameters based on identified operating conditions. The operating characteristics of the power generation process vary significantly with changing operating conditions. For example, startup conditions exhibit high thermal inertia, peak-shaving conditions experience drastic load changes, and combustion instability conditions result in significant temperature fluctuations. Traditional fixed-parameter models struggle to adapt to these changing conditions. The innovation of this method lies in its design of 13 operating condition identifiers (including 2 CFB-specific conditions), capable of online identification of the current operating condition type and automatically loading the corresponding model parameter configuration. Operating condition identification employs a multi-index fusion method, comprehensively considering indicators such as load change rate (% / min), temperature change rate (°C / min), and pressure change rate (MPa / min). Each operating condition corresponds to an independent subset of parameters, including the number of attention heads, latent variable dimensions, and learning rate for each Transformer layer. This adaptive operating condition mechanism enables the model to flexibly respond to various operating states of the power generation process, improving prediction accuracy by 31.7% to 45.7% under various operating conditions.

[0120] Fifth, the distributed collaborative incremental learning and update mechanism: This method proposes a distributed, collaborative incremental learning and update mechanism to address the challenge of continuous optimization after model deployment. Power generation equipment faces aging and changing operating conditions during long-term operation, requiring continuous model updates to maintain prediction accuracy. The innovation of this method lies in its complete incremental update process, comprising three stages: experience replay, knowledge distillation, and concept drift detection. The experience replay mechanism prevents model forgetting by maintaining a balance between old and new data (50%:50% sampling ratio); the knowledge distillation mechanism maintains stability by transferring knowledge from the old model to the new model; and the concept drift detection mechanism promptly identifies changes in data distribution and triggers comprehensive retraining through statistical testing (KS test, significance level p<0.05). The distributed architecture enables efficient incremental updates without impacting online services. Experiments show that, after adopting the incremental learning and update mechanism, the model's prediction accuracy decay is controlled within 5% over a 12-month deployment period.

[0121] The time series prediction method provided in the above embodiments was specifically verified through experiments, and the comparison results are shown in Tables 1 and 2: Table 1 Comparison of Main Steam Temperature Prediction Accuracy

[0122] The above data is based on the test results of a 600MW thermal power unit main steam temperature prediction mission, and the test data includes four quarters of operational data. The physical consistency score ranges from 0 to 1, with a higher score indicating that the prediction result is more consistent with physical laws.

[0123] Table 2 compares the predictive performance under different operating conditions.

[0124] The above data compares the main steam temperature prediction accuracy (MAE) of the method in this embodiment with that of other comparative methods under different operating conditions. The test conditions cover 11 typical operating conditions of thermal power units, and this prediction method achieves the best or second-best prediction accuracy under all conditions.

[0125] The timing prediction method provided in this embodiment will be further implemented as follows: A 600MW supercritical thermal power unit in a power plant adopted this method for multi-step prediction of main steam temperature. The unit is equipped with a complete distributed control system (DCS) capable of real-time acquisition of key parameters such as main steam temperature (°C), main steam pressure (MPa), feedwater temperature (°C), flue gas temperature (°C), and load command (MW). The data acquisition frequency is 1 second / time, and the historical data storage period is 1 year. Specific hardware and software configuration parameters are shown in Tables 3 and 4. Table 3 shows the hardware configuration parameters.

[0126] Table 4 shows the software configuration parameters.

[0127] Based on the aforementioned hardware and software configuration, the specific testing steps are as follows: First, the collected raw data is preprocessed, including noise reduction, missing value imputation, and standardization. The data preprocessing module automatically identified and processed approximately 0.3% of outlier data points and 0.8% of missing data points. Then, a multi-scale time series decomposition module is used to decompose the main steam temperature sequence into a trend component (reflecting the long-term aging trend of the unit), a periodic component (reflecting the day-night load cycle), and a random component (reflecting load fluctuations and disturbances).

[0128] During the model training phase, a hierarchical latent variable encoder is used to encode the three components into their respective latent variable spaces. The trend latent variable dimension is set to 64, the periodic latent variable dimension to 32, and the random latent variable dimension to 64. The physical constraint injection module embeds the boiler water-cooled wall heat transfer model and the superheater heat transfer model into the training process to ensure that the prediction results satisfy the law of conservation of energy.

[0129] After training, the model made 24-hour advance predictions of the main steam temperature. Under stable operation at rated load, the predicted MAE was 0.28°C, RMSE was 0.38°C, the 90% prediction range coverage reached 91.5%, and the physical consistency score was 0.98. Under variable load operation (load change rate ±5% / min), the MAE was 0.33°C, a 40% reduction compared to the baseline method. Under deep peak shaving operation (40% load), the MAE was 0.38°C, a 43.3% reduction compared to the baseline method.

[0130] During operation, the prediction results are displayed to operators in real time through a visualization module. When a prediction deviation exceeding 3°C is detected, the system automatically issues a level-two alarm and displays a possible cause analysis. Six months of actual operation show that after adopting this method, the number of main fuel temperature adjustment operations caused by temperature prediction deviations has decreased by approximately 35%, and the unit's operational stability and economy have been significantly improved.

[0131] Based on the above tests, the main steam temperature prediction test data are shown in Table 5: Table 5 shows the main steam temperature prediction test data.

[0132] like Figure 3 As shown, specifically as follows Figure 3As shown in (a), (b), (c), and (d) of this embodiment, the prediction method adopts the LST-SAM three-component Transformer architecture and is compared with the LSTM single encoder, the Informer long sequence prediction model, and the Autoformer autoregressive model. The evaluation metrics include MAE (unit: °C), RMSE (unit: °C), and 90% prediction interval coverage (PI, unit: %). The test data comes from 11 typical operating conditions of a 600MW unit (start-up / shutdown / deep peak shaving / load increase / decrease, etc.), with ≥5000 test samples and a sampling period of 5 seconds / time. Figure 3 The data is presented in a grouped bar chart format, with the horizontal axis representing 11 operating conditions and the vertical axis representing the MAE value (°C). Each group of four bars represents LST-SAM / LSTM / Informer / Autoformer, distinguished by dark blue / light blue / orange / gray, and the name of the optimal method for each operating condition is labeled below. Figure 3 This study is the first to compare multi-dimensional benchmark methods under 11 operating conditions in the thermal power field, and includes a 90% confidence interval coverage assessment (rather than a single deterministic indicator). Performance comparisons are made on the main steam temperature prediction task for 600MW thermal power units. (a) MAE comparison: our method's MAE is 0.31°C, significantly lower than all benchmark methods; (b) RMSE comparison: our method's RMSE is 0.42°C; (c) 90% prediction interval coverage comparison: our method reaches 91.2%; (d) Physical consistency score comparison: our method's score reaches 0.97, far exceeding all benchmark methods.

[0133] For example, this embodiment also provides a specific implementation of the above-mentioned time series prediction method, and the implementation process is as follows: A 350MW circulating fluidized bed (CFB) boiler unit in a power plant uses this method for bed temperature prediction and optimized control. CFB boilers are characterized by uneven bed temperature distribution and large combustion fluctuations, making it difficult for traditional prediction methods to achieve satisfactory results. Specific hardware and software configuration parameters are shown in Tables 6 and 7. Table 6 shows the hardware configuration parameters.

[0134] Table 7 shows the software configuration parameters.

[0135] To address the unique characteristics of CFB boilers, this method is specifically optimized based on the standard architecture. First, the operating condition identification module is expanded by adding two sub-modules: fluidization state identification and return material circulation state identification, bringing the total number of operating condition identification categories to 13. Second, the physical constraint injection module incorporates fluidization kinetic constraints and material balance constraints unique to CFB boilers. Fluidization kinetic constraints ensure that the predicted bed temperature matches parameters such as fluidization velocity and material concentration; material balance constraints ensure the mass balance between limestone desulfurization and fuel combustion reactions.

[0136] The model training data includes the unit's historical operating data over the past three years, totaling approximately 26,000 hours of continuous operation records. A distributed training architecture was used during training, with four computing nodes working in parallel, and the total training time was approximately eight hours.

[0137] Post-deployment predictive performance test results: Under typical operating conditions, the predicted bed temperature MAE is 2.8°C, a 33.3% reduction compared to the LSTM baseline method (MAE=4.2°C). Under conditions of fluctuating fuel quality (coal calorific value change ±15%), the MAE of this method is 3.2°C, compared to 5.1°C for the baseline method, a 37.3% reduction. Under abnormal fluidization conditions, the physical consistency score of this method reaches 0.91, significantly higher than the baseline method's 0.64. Specific CFB bed temperature prediction test data are shown in Table 8. Table 8 shows the CFB bed temperature prediction test data.

[0138] Based on the bed temperature prediction results, the control system enabled advance adjustment of the bed temperature, reducing the bed temperature fluctuation range from ±15°C to ±8°C, improving desulfurization efficiency by approximately 2 percentage points, and saving approximately 800 tons of limestone annually. Simultaneously, the number of unplanned shutdowns due to abnormal bed temperatures decreased from an average of 3 times per year to 1 time, significantly improving unit availability.

[0139] For example, such as Figure 5 As shown, this embodiment also provides a time-series prediction system for power generation processes across multiple time scales, including: The decomposition module is used to decompose the raw time-series data of real-time acquisition of multiple operating parameters into several intrinsic mode functions based on the empirical mode decomposition method. The intrinsic mode functions are then merged into random components, periodic components and trend components by wavelet transform according to the periodic scale, while extracting the statistical features and time-frequency features of each component. The mapping module is used to construct a three-layer latent variable encoder that corresponds to the trend component, periodic component, and random component, respectively. The features of the corresponding components are input into each layer encoder for feature mapping to obtain the trend latent variable, periodic latent variable, and random latent variable. The calculation module is used to construct the thermodynamic physical model and mass balance model of the power generation process, transform the physical model into a differentiable physical constraint layer, and use the latent variables of each layer and the original observation data as inputs to the physical constraint layer to calculate and generate the physical consistency loss term. The inference module is used to perform time-series feature inference by inputting the trend latent variable, periodic latent variable and random latent variable into the corresponding Transformer sub-module based on the physical consistency loss term, and obtain multi-scale fusion inference features. The prediction module is used to input multi-scale fused inference features into the point prediction output head, interval prediction output head and confidence prediction output head set in parallel. The three prediction tasks are jointly optimized through a multi-task loss function, and the output includes the point prediction value, prediction interval and confidence score. The correction module is used to perform physical consistency verification on the initial prediction results based on preset physical constraint rules. For prediction values ​​that fail the verification, iterative correction is performed along the reverse direction of the physical constraint gradient to obtain the final prediction result.

[0140] The present invention also provides a time-series prediction device for a power generation process with multiple time scales, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the time-series prediction method for a power generation process with multiple time scales.

[0141] The present invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the multi-timescale timing prediction method for the power generation process.

[0142] When the processor executes the computer program, it implements the steps of multi-timescale time-series prediction of the above-mentioned power generation process. For example, based on the empirical mode decomposition method, the original time-series data of real-time acquisition of multiple types of operating parameters are decomposed into several intrinsic mode functions. The intrinsic mode functions are merged into random components, periodic components and trend components according to the periodic scale by wavelet transform, and the statistical features and time-frequency features of each component are extracted at the same time. A three-layer latent variable encoder is constructed, corresponding to the trend component, periodic component, and random component respectively. The features of the corresponding components are input into each layer encoder for feature mapping to obtain the trend latent variable, periodic latent variable, and random latent variable. A thermodynamic physical model and a mass balance model of the power generation process are constructed. The physical model is transformed into a differentiable physical constraint layer. The latent variables of each layer and the original observation data are used as inputs to the physical constraint layer to calculate and generate the physical consistency loss term. Based on the physical consistency loss term, the trend latent variable, periodic latent variable and random latent variable are respectively input into the corresponding Transformer sub-module for time series feature inference, and multi-scale fusion inference features are obtained. The multi-scale fusion inference features are input into the point prediction output head, interval prediction output head and confidence prediction output head set in parallel. The three prediction tasks are jointly optimized through a multi-task loss function, and the initial prediction results including point prediction values, prediction intervals and confidence scores are output. The initial prediction results are physically consistent with the preset physical constraint rules. For predictions that fail the check, the results are iteratively corrected in the opposite direction of the physical constraint gradient to obtain the final prediction results.

[0143] For example, the computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing preset functions, the instruction segments describing the execution process of the computer program in the multi-timescale timing prediction device for the power generation process. For example, the computer program can be divided into a decomposition module, a mapping module, a calculation module, an inference module, a prediction module, and a correction module; its specific functions are as follows: The decomposition module decomposes the raw time-series data of multiple operating parameters acquired in real time into several intrinsic mode functions (IMFs) based on the empirical mode decomposition method. It then merges these IMFs into random components, periodic components, and trend components using wavelet transform at a periodic scale, while simultaneously extracting the statistical and time-frequency features of each component. The mapping module constructs three-layer latent variable encoders corresponding to the trend component, periodic component, and random component, respectively. It inputs the features of the corresponding components into each encoder layer for feature mapping to obtain the trend latent variable, periodic latent variable, and random latent variable. The calculation module constructs a thermodynamic physical model and a mass balance model of the power generation process, transforming the physical model into a differentiable physical constraint layer. Using latent variables at each layer and the original observation data as inputs to the physical constraint layer, a physical consistency loss term is calculated and generated. The inference module, based on the physical consistency loss term, inputs trend latent variables, periodic latent variables, and random latent variables into their respective Transformer sub-modules for time-series feature inference, obtaining multi-scale fusion inference features. The prediction module inputs the multi-scale fusion inference features into parallel point prediction output heads, interval prediction output heads, and confidence prediction output heads, jointly optimizing the three prediction tasks through a multi-task loss function, outputting initial prediction results including point prediction values, prediction intervals, and confidence scores. The correction module performs physical consistency verification on the initial prediction results based on preset physical constraint rules, iteratively correcting prediction values ​​that fail the verification along the reverse direction of the physical constraint gradient to obtain the final prediction result.

[0144] The multi-timescale timing prediction device for the power generation process can be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. This device may include, but is not limited to, processors and memory. Those skilled in the art will understand that the above are examples of multi-timescale timing prediction devices for the power generation process and do not constitute a limitation on such devices. The device may include more components than described above, or combine certain components, or use different components. For example, the multi-timescale timing prediction device may also include input / output devices, network access devices, buses, etc.

[0145] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or any conventional processor. The processor is the control center for the multi-timescale timing prediction of the power generation process, connecting various parts of the multi-timescale timing prediction equipment for the entire power generation process via various interfaces and lines.

[0146] The memory can be used to store the computer program and / or modules. The processor realizes various functions of the multi-timescale time-series prediction device for the power generation process by running or executing the computer program and / or modules stored in the memory and calling the data stored in the memory.

[0147] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function (such as sound playback, image playback, etc.). The data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). Furthermore, the memory may include high-speed random access memory and non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital cards (SD cards), flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0148] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned time-series prediction method for a power generation process across multiple time scales.

[0149] If the modules / units of the multi-timescale time-series prediction system for the power generation process are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0150] Based on this understanding, the present invention can implement all or part of the processes in the multi-timescale time-series prediction method for power generation, which can also be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the multi-timescale time-series prediction method for power generation. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or a preset intermediate form, etc.

[0151] The computer-readable storage medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0152] It should be noted that the content contained in the computer-readable storage medium may be appropriately increased or decreased according to the requirements of legislation and methodological practice in the jurisdiction. For example, in some jurisdictions, according to legislation and methodological practice, the computer-readable storage medium may not include electrical carrier signals and telecommunication signals.

[0153] Compared with existing prediction methods, this invention provides a time-series prediction method and related apparatus for power generation processes across multiple time scales, which has the following advantages: This invention first collects and preprocesses the operating parameters of power generation equipment in real time. Then, it uses empirical mode decomposition and wavelet transform to decompose the time-series signal into three levels: trend, periodic, and random. A three-layer latent variable encoder then maps each component to its corresponding latent variable space. During model training, a physical constraint injection module embeds prior physical knowledge such as thermodynamic laws and mass balance equations in a differentiable manner to ensure that the prediction results meet physical consistency constraints. A hierarchical Transformer architecture is used to process the latent variables at each level, enabling time-series inference across multiple time scales. This method simultaneously outputs the results of three tasks: point prediction values, prediction intervals, and confidence scores, providing operators with richer decision support information. This invention achieves a MAE accuracy of 0.31°C on the main steam temperature prediction task of a 600MW supercritical thermal power unit, significantly outperforming existing methods, and the prediction results exhibit better physical consistency, effectively supporting optimized operation and intelligent control of the power generation process.

[0154] The above embodiments are merely one of the implementation methods for achieving the technical solution of the present invention. The scope of protection claimed by the present invention is not limited to this embodiment, but also includes any variations, substitutions and other implementation methods that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the present invention.

Claims

1. A time-series prediction method for a power generation process across multiple time scales, characterized in that, include: Based on the empirical mode decomposition method, the original time series data of multiple operating parameters acquired in real time are decomposed into several intrinsic mode functions. By wavelet transform according to the periodic scale, each intrinsic mode function is merged into random component, periodic component and trend component. At the same time, the statistical features and time-frequency features of each component are extracted. A three-layer latent variable encoder is constructed, corresponding to the trend component, periodic component, and random component respectively. The features of the corresponding components are input into each layer encoder for feature mapping to obtain the trend latent variable, periodic latent variable, and random latent variable. A thermodynamic physical model and a mass balance model of the power generation process are constructed. The physical model is transformed into a differentiable physical constraint layer. The latent variables of each layer and the original observation data are used as inputs to the physical constraint layer to calculate and generate the physical consistency loss term. Based on the physical consistency loss term, the trend latent variable, periodic latent variable and random latent variable are respectively input into the corresponding Transformer sub-module for time series feature inference, and multi-scale fusion inference features are obtained. The multi-scale fusion inference features are input into the point prediction output head, interval prediction output head and confidence prediction output head set in parallel. The three prediction tasks are jointly optimized through a multi-task loss function, and the initial prediction results including point prediction values, prediction intervals and confidence scores are output. The initial prediction results are physically consistent with the preset physical constraint rules. For predictions that fail the check, the results are iteratively corrected in the opposite direction of the physical constraint gradient to obtain the final prediction results.

2. The time-series prediction method for a power generation process with multiple time scales according to claim 1, characterized in that, The process of acquiring the raw time-series data of the various operating parameters includes: acquiring operating parameters such as main steam temperature, main steam pressure, feedwater temperature, flue gas temperature, load command, fuel flow rate, and valve opening degree in real time from the distributed control system of the thermal power unit; performing quality control on the raw time-series data after acquisition, using the three-standard-deviation criterion to identify outliers exceeding the threshold range, identifying missing data segments with a continuous missing duration exceeding a preset time, and identifying high-frequency oscillation noise data in the time series; correcting outlier data using an interpolation method based on a variational autoencoder, filling missing data using a filling method based on latent variable inference, and using the processed quality-controlled time-series data for empirical mode decomposition.

3. The time-series prediction method for a power generation process at multiple time scales according to claim 1, characterized in that, The process of merging the intrinsic mode functions into random components, periodic components, and trend components using wavelet transform on a periodic scale is as follows: The decomposition layer is set to 3 layers, where the period of the high-frequency random component is less than 1 hour, the period of the mid-frequency periodic component is 1 to 24 hours, and the period of the low-frequency trend component is greater than 24 hours; the wavelet transform uses the db4 wavelet basis function; the extracted statistical features include mean, variance, skewness, and kurtosis; the extracted time-frequency features include instantaneous frequency and instantaneous amplitude; the extracted component features are input into the corresponding three-layer latent variable encoder.

4. The time-series prediction method for a power generation process with multiple time scales according to claim 1, characterized in that, The three-layer latent variable encoders are a trend latent variable encoder, a periodic latent variable encoder, and a random latent variable encoder. The trend latent variable encoder uses a one-dimensional convolutional network to extract trend features, which are then mapped through a fully connected layer to obtain 64-dimensional trend latent variables. The periodic latent variable encoder uses a Fourier transform module to extract periodic features, which are then mapped through a fully connected layer to obtain 32-dimensional periodic latent variables. The random latent variable encoder uses a variational autoencoder framework to encode 64-dimensional random latent variables. The Kolb-Leibler divergence of the latent variables output by each encoder layer is used as a regularization term in model training. The three types of latent variables generated are input into the corresponding Transformer submodules and the physical constraint layer, respectively.

5. The time-series prediction method for a power generation process at multiple time scales according to claim 1, characterized in that, The thermodynamic physical model includes a boiler water-cooled wall heat transfer model, a superheater heat transfer model, and a steam state equation. Physical constraint terms are added to the loss function using the Lagrange multiplier method, and the weights of the physical constraints are dynamically adjusted through learnable parameters. The calculated physical consistency loss term is added to the multi-task loss function and participates in the model training and optimization process together with the prediction task loss.

6. The time-series prediction method for a power generation process at multiple time scales according to claim 1, characterized in that, The Transformer submodules are a Trend Transformer submodule, a Periodic Transformer submodule, and a Random Transformer submodule. The Trend Transformer submodule employs a 12-layer full attention structure with 8 attention heads per layer and a latent variable dimension of 512. The Periodic Transformer submodule employs a 6-layer periodic-aware sparse attention structure with 8 attention heads per layer and a latent variable dimension of 256. The Random Transformer submodule employs a 6-layer standard self-attention structure with 8 attention heads per layer and a latent variable dimension of 512. The three submodules interact with each other through a cross-layer attention mechanism. The weights of the cross-layer attention are determined by a learnable parameter matrix, and the resulting submodules are fused to generate multi-scale fusion inference features.

7. The time-series prediction method for a power generation process at multiple time scales according to claim 1, characterized in that, The point prediction output head uses a two-layer fully connected network to output the predicted value for the next 24 hours; the interval prediction output head uses a quantile regression method to output prediction intervals for five quantiles: 10%, 25%, 50%, 75%, and 90%; the confidence output head uses a Bayesian linear regression layer to output the prediction variance estimate; the three prediction tasks are jointly optimized through a multi-task loss function with dynamic weights. The physical consistency verification includes temperature range check, pressure range check, thermodynamic relationship check, and mass balance check. Predicted values ​​that fail the verification are iteratively corrected with a preset correction step size, and the final prediction result is obtained and output after correction.

8. A time-series prediction system for a power generation process with multiple time scales, characterized in that, include: The decomposition module is used to decompose the raw time-series data of real-time acquisition of multiple operating parameters into several intrinsic mode functions based on the empirical mode decomposition method. The intrinsic mode functions are then merged into random components, periodic components and trend components by wavelet transform according to the periodic scale, while extracting the statistical features and time-frequency features of each component. The mapping module is used to construct a three-layer latent variable encoder that corresponds to the trend component, periodic component, and random component, respectively. The features of the corresponding components are input into each layer encoder for feature mapping to obtain the trend latent variable, periodic latent variable, and random latent variable. The calculation module is used to construct the thermodynamic physical model and mass balance model of the power generation process, transform the physical model into a differentiable physical constraint layer, and use the latent variables of each layer and the original observation data as inputs to the physical constraint layer to calculate and generate the physical consistency loss term. The inference module is used to perform time-series feature inference by inputting the trend latent variable, periodic latent variable and random latent variable into the corresponding Transformer sub-module based on the physical consistency loss term, and obtain multi-scale fusion inference features. The prediction module is used to input multi-scale fused inference features into the point prediction output head, interval prediction output head and confidence prediction output head set in parallel. The three prediction tasks are jointly optimized through a multi-task loss function, and the output includes the point prediction value, prediction interval and confidence score. The correction module is used to perform physical consistency verification on the initial prediction results based on preset physical constraint rules. For prediction values ​​that fail the verification, iterative correction is performed along the reverse direction of the physical constraint gradient to obtain the final prediction result.

9. A time-series prediction device for a power generation process across multiple time scales, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the multi-timescale time-series prediction method for the power generation process according to any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it is used to implement the multi-timescale time-series prediction method for the power generation process according to any one of claims 1-7.

Images