Industrial equipment state quantity long time domain prediction method and device

Abstract

The application provides an industrial equipment state quantity long-time domain prediction method and device, and relates to the technical field of industrial intelligent monitoring. The optimal lag and response strength of each variable to the target variable are explicitly identified through dynamic response lag analysis, the physical transmission characteristics of the industrial system are accurately described, and the model is interpretable. The sparse graph prior matrix constructed based on the response strength and the optimal lag is injected into the attention mechanism as a soft constraint, the attention learning is more stable and robust, and the noise sensitivity is reduced. The trend, dynamic response and residual are treated separately through three-component decoupling prediction, the modeling burden is reduced, and the long-time domain prediction accuracy is improved. The prediction uncertainty is learned and the confidence is output through joint learning, the error accumulation is effectively inhibited through the online correction mechanism, and the long-term prediction reliability is improved. The model effective service cycle is prolonged through rolling calibration to automatically adapt to the state distribution drift of the equipment.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent monitoring technology, and in particular to a method and apparatus for long-term prediction of the state variables of industrial equipment. Background Technology

[0002] During long-term operation, industrial equipment collects a large amount of multi-source, multi-frequency, and multi-dimensional status data through systems such as SCADA and DCS, forming high-dimensional multivariate time series. Long-term time-domain prediction of the future state of equipment based on historical multivariate time series data is a key foundation for achieving predictive maintenance and operational optimization.

[0003] Existing time series forecasting methods mainly include statistical autoregressive models, recurrent neural networks, temporal convolutional networks, and Transformer-like attention models. However, in real-world industrial long-term forecasting scenarios, the following problems still exist: the dynamic response lag relationship between variables is not explicitly modeled, leading to the loss of physical transfer characteristics; the attention mechanism is entirely data-driven, which is prone to generating spurious correlations under industrial noise and operating condition changes, resulting in unstable relationship learning; long-term multi-step forecasting errors accumulate significantly; there is a lack of decomposition mechanisms for trends, responses, and disturbances, leading to ineffective consumption of model capacity; and after model deployment, there is a lack of online confidence assessment and self-correction capabilities, making it impossible to dynamically correct for high-risk moments. Summary of the Invention

[0004] The purpose of this invention is to provide a method and apparatus for long-term prediction of state variables of industrial equipment, which can accurately characterize the physical transmission characteristics of industrial systems, reduce noise sensitivity, reduce modeling burden, improve long-term prediction accuracy, improve long-term prediction reliability, and extend the effective service life of the model.

[0005] In a first aspect, the present invention provides a long-term time-domain prediction method for state variables of industrial equipment, comprising: Historical multi-source time series data of industrial equipment are acquired and preprocessed. Key driving variables are selected from candidate variables through dynamic response lag analysis, and a sparse graph prior matrix is ​​constructed based on the response intensity and optimal lag of the key driving variables. Based on the historical sequences of key driving variables and the prior matrix of a sparse graph, the basic prediction model is trained until the preset training completion conditions are met to obtain the prediction model. The basic prediction model includes: temporal encoding, graph prior-constrained attention, and three-component decoupled prediction. The loss function of the basic prediction model is based on prediction uncertainty, mean square error, three-component orthogonalization loss, and weight regularization joint constraint. Real-time running data is preprocessed and then input into the trained prediction model to obtain the prediction sequence and the confidence level of each step in the prediction sequence. Online correction is performed when the confidence level is less than the preset confidence level threshold. Rolling calibration is performed based on the recent prediction performance after the prediction model is deployed.

[0006] In some preferred embodiments of the present invention, the sampling reference time is predetermined; preprocessing includes: Candidate variables are interpolated and aligned at the sampling reference time. Based on the preset missing rate threshold, missing data are processed by interpolation filling, interpolation of the mean under the same working conditions, or direct removal. Based on the segmented statistics of working conditions, outliers that deviate from the mean by more than a preset standard deviation are replaced by the neighborhood median. Candidate variables are normalized using the statistics of the training set.

[0007] In some preferred embodiments of the present invention, a lag search set is pre-defined; key driving variables are selected from candidate variables through dynamic response lag analysis, including: Each lag candidate variable sequence within the lag search set is shifted and paired with the target variable sequence. The absolute value of the Pearson correlation coefficient and the normalized mutual information based on histogram binning are calculated respectively. The absolute value and normalized mutual information are weighted and fused by preset weights to obtain the composite response intensity. The lag that maximizes the composite response strength is taken as the optimal lag of the candidate variable, and the maximum composite response strength is taken as the response strength of the candidate variable. The mean of the response intensity of all candidate variables is used as the adaptive threshold, and variables with response intensity greater than the adaptive threshold are retained as key driving variables.

[0008] In some preferred embodiments of the present invention, a sparse graph prior matrix is ​​constructed based on the response intensity and optimal lag of key driving variables, including: Using the target variable as the central node, the original weights are obtained by normalizing the response strengths of each key driving variable. The original weights are corrected based on the exponential decay factor associated with the optimal lag. The corrected weights are used as the edge weights of the connections between variables. Connections with edge weights less than a preset pruning threshold are removed to obtain the sparse graph prior matrix. During model training, the sparse graph prior matrix is ​​updated once every preset training period based on the statistics of the current training set.

[0009] In some preferred embodiments of the present invention, timing coding includes: Extract the historical observation sequence of each key driving variable, input it into a temporal coding module with a causal convolution, dilated causal convolution, temporal convolutional network or recurrent neural network structure, and generate a high-order representation vector; All high-order representation vectors are stacked into a multivariate feature representation matrix, and different key driving variables share the same set of encoding parameters.

[0010] In some preferred embodiments of the present invention, graph-prior constraint attention includes: The multivariate feature representation matrix is ​​mapped to a query matrix, a key matrix, and a value matrix through a trainable projection. The sparse graph prior matrix is ​​injected into the attention in logarithmic form, an adaptive sparsification strategy based on the number of variables is applied, and the retained attention connections are renormalized. The attention output is obtained by weighted summation of the value matrix using the renormalized attention weights, stacking multiple layers, and introducing residual connections and layer normalization between layers.

[0011] In some preferred embodiments of the present invention, the three-component decoupled prediction includes: A moving average is applied to the historical sequence of the target variable, and the resulting trend sequence is linearly extrapolated to generate trend components. The attention output of the graph prior constraint attention output is flattened and then input into a multilayer perceptron to map and generate dynamic response components. The historical residual sequence is input into a gated recurrent unit network, and residual components are generated through linear mapping; wherein, the historical residual sequence is obtained by subtracting the estimated values ​​of the historical parts corresponding to the trend component and the dynamic response component from the historical sequence of the target variable; The trend component, dynamic response component, and residual component are summed to form the prediction sequence.

[0012] In some preferred embodiments of the present invention, the loss function of the basic prediction model is based on a joint constraint of prediction uncertainty, mean square error, three-component orthogonalization loss, and weight regularization, including: Set up an independent branch to output variance correlation in the prediction header of the dynamic response component; The negative log-likelihood loss is calculated based on the variance correlation, and the negative log-likelihood loss, mean squared error loss, three-component orthogonalization loss, and weighted regularization term are weighted and summed as the joint loss function of the basic prediction model.

[0013] In some preferred embodiments of the present invention, online calibration includes: The confidence level of each step in the prediction sequence is calculated based on the variance correlation quantity output by the prediction model, so that the confidence level decreases monotonically as the uncertainty increases. When the confidence level is less than the preset confidence level threshold, the regularized linear model trained based on recent measured values ​​and the previously generated prediction values ​​in the prediction sequence is weighted and fused with the prediction mean of the prediction sequence to obtain the corrected prediction value. Rolling calibration, including: After the prediction model is deployed, at preset time intervals, the calibration score and empirical coverage of the prediction distribution are calculated based on recent measured values, historical predicted values ​​and variance correlation. When the empirical coverage deviates from the preset ideal coverage by more than the preset allowable deviation, the confidence base threshold is adjusted within the protection interval by a preset step size.

[0014] Secondly, the present invention provides a long-term time-domain prediction device for industrial equipment state variables, comprising: The data processing module is used to acquire and preprocess historical multi-source time-series data of industrial equipment, select key driving variables from candidate variables through dynamic response lag analysis, and construct a sparse graph prior matrix based on the response strength and optimal lag of the key driving variables. The model training module is used to train the basic prediction model based on the historical sequences of key driving variables and the sparse graph prior matrix until the preset training completion conditions are met, thus obtaining the prediction model. The basic prediction model includes: temporal encoding, graph prior constrained attention, and three-component decoupled prediction. The loss function of the basic prediction model is based on prediction uncertainty, mean square error, three-component orthogonalization loss, and weight regularization joint constraint. The data prediction module is used to input the preprocessed real-time running data into the trained prediction model to obtain the prediction sequence and the confidence level of each step in the prediction sequence, and to perform online correction when the confidence level is less than the preset confidence level threshold; wherein, after the prediction model is deployed, rolling calibration is performed based on the recent prediction performance.

[0015] This invention brings the following beneficial effects: This invention provides a method and apparatus for long-term prediction of state variables of industrial equipment. It explicitly identifies the optimal lag and response strength of each variable to the target variable through dynamic response lag analysis, accurately characterizing the physical transfer characteristics of the industrial system and making the model interpretable. By injecting a sparse graph prior matrix constructed based on response strength and optimal lag as a soft constraint into the attention mechanism, attention learning becomes more stable and robust, reducing noise sensitivity. Three-component decoupled prediction separates and conquers the trend, dynamic response, and residuals, reducing the modeling burden and improving long-term prediction accuracy. Joint learning predicts uncertainty and outputs confidence scores, with an online correction mechanism effectively suppressing error accumulation and improving long-term prediction reliability. Rolling calibration automatically adapts to equipment state distribution drift, extending the effective service life of the model. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 A flowchart of a long-term prediction method for state variables of industrial equipment provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the overall RMSE comparison results of multiple prediction methods when the prediction time domain is 60, provided by an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the overall MAE comparison results of multiple prediction methods when the prediction time domain is 60, provided by an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the overall RMSE comparison results of multiple prediction methods when the prediction time domain is 90 seconds, provided by an embodiment of the present invention. Figure 5 This is a schematic diagram illustrating the overall MAE comparison results of multiple prediction methods when the prediction time domain is 90 seconds, provided by an embodiment of the present invention. Figure 6 This is a schematic diagram of the structure of a long-term prediction device for the state variables of industrial equipment provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0018] Icons: 310 - Data processing module; 320 - Model training module; 330 - Data prediction module; 400 - Memory; 401 - Processor; 402 - Bus; 403 - Communication interface. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0020] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0021] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0022] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. In addition, the terms "first," "second," "third," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0023] Furthermore, terms such as "horizontal," "vertical," and "sag" do not imply that components must be absolutely horizontal or suspended, but rather that they can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal relative to "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0024] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0025] During long-term operation, industrial equipment continuously collects a large amount of multi-source, multi-frequency, and multi-dimensional status data through SCADA systems, DCS systems, and distributed intelligent sensor networks, forming high-dimensional multivariate time series. Long-term time-domain prediction of the equipment's future state based on historical multivariate time-series data (long-term time-domain prediction typically refers to multi-step forward prediction with a prediction step size of tens to hundreds of steps, such as scenarios with H≥60 or H≥96) is a key foundation for achieving predictive maintenance, operating parameter optimization, and early fault warning.

[0026] Currently, mainstream time series prediction techniques can be roughly divided into four categories: (1) statistical autoregressive models represented by ARIMA and VAR; (2) recurrent neural network models represented by LSTM and GRU; (3) temporal convolutional models represented by TCN; and (4) Transformer-type attention prediction models represented by Informer, Autoformer, PatchTST, and iTransformer. These methods have made some progress on standard benchmark datasets, but in real industrial long-term prediction scenarios, they still have the following deep-seated technical defects: (1) The dynamic response lag relationship between variables is not explicitly modeled. Industrial systems have typical physical transmission characteristics. For example, in thermal power boilers, after the feedwater flow rate changes, it needs to go through a series of physical links such as feedwater pipelines, economizers, and steam-water separators before it is finally reflected as a change in the steam drum water level. The whole process usually has a delay of tens of seconds to several minutes. In wind turbine units, the pitch angle adjustment signal often spans dozens of sampling cycles from the time the controller sends it to the time the main bearing temperature responds. Existing Transformer class methods construct Query and Key based on the same time segment, and the coupling between variables is "instantaneous" by default, which cannot characterize causal transmission across time segments.

[0027] (2) The attention mechanism is entirely data-driven, and the relationship learning is unstable. The standard self-attention weights are completely fitted by the data. In industrial scenarios, the sensor noise is high, the operating conditions change frequently, and the effective samples are sparse. The attention weights are prone to "spurious associations", and the performance drops sharply when the operating conditions drift.

[0028] (3) Long-term multi-step prediction errors accumulate severely. Direct multi-step prediction is difficult to simultaneously take into account both low-frequency trends and high-frequency fluctuations; in recursive multi-step prediction, the single-step error is gradually amplified, and the RMSE usually increases by 30% to 80% for every doubling of the prediction step length.

[0029] (4) Lack of decomposition mechanism for trend, response and disturbance. Industrial signals are essentially composed of slow-changing trends, dynamic responses driven by other variables and unexplainable random disturbances. Existing methods couple the three into the same network for end-to-end learning, which wastes the model capacity.

[0030] (5) Lack of online confidence assessment and self-correction capabilities. After the model is deployed, a fixed prediction strategy is adopted, which cannot dynamically correct for high-risk moments based on prediction uncertainty.

[0031] In summary, there is an urgent need to propose a long-term prediction method that can explicitly model the dynamic response of variables, introduce structural prior constraints, decouple multiple components, and has online self-correction capabilities.

[0032] This invention aims to provide a long-term prediction method for industrial equipment state variables based on dynamic response decoupling and graph prior constraint attention mechanism. It solves five major problems existing in the prior art for long-term prediction in industrial applications: "response lag not modeled, attention unstable, significant error accumulation, multi-component not decoupled, and no confidence correction". Thus, it significantly improves prediction accuracy, stability and interpretability in actual industrial environments with complex operating conditions and high noise.

[0033] The following detailed description of some embodiments of the present invention is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0034] This invention provides a long-term time-domain prediction method for state variables of industrial equipment, see [link to relevant documentation]. Figure 1 The flowchart shown in this embodiment of the invention provides a long-term prediction method for state variables of industrial equipment. The method includes: Step S102: Obtain historical multi-source time series data of industrial equipment and preprocess it. Select key driving variables from candidate variables through dynamic response lag analysis, and construct a sparse graph prior matrix based on the response intensity and optimal lag of the key driving variables.

[0035] Specifically, this step is the data preparation and structured prior construction stage for long-term time-domain prediction of industrial equipment state variables. Its purpose is to transform the original multivariate heterogeneous time-series data collected by systems such as SCADA and DCS into standardized historical window inputs. Simultaneously, it utilizes statistical and information theory methods to extract the physical transit lag relationships between variables and encodes these relationships into a sparse graph prior matrix that can be incorporated into subsequent deep models. This stage includes three core components: data preprocessing, dynamic response lag analysis, and sparse graph prior matrix construction.

[0036] The data preprocessing stage is responsible for time alignment, missing value imputation, outlier cleaning, and dimensional normalization of the raw industrial data, eliminating inconsistencies and noise interference from multi-source data, and forming a unified historical window input matrix to provide a high-quality data foundation for subsequent analysis. The dynamic response lag analysis stage simultaneously calculates the Pearson correlation coefficient (linear measure) and normalized mutual information (nonlinear measure) of each candidate variable and the target variable at different time lags, and obtains the composite response strength through weighted fusion. This systematically identifies the optimal response lag time and comprehensive dependence strength of each variable on the target variable. Then, an adaptive threshold is used to filter out the set of key variables with significant driving effects, explicitly encoding the inherent physical transmission delay characteristics of the industrial system into quantifiable lag parameters. The sparse graph prior matrix construction step transforms the aforementioned statistical conclusions into a directed weighted graph structure. Using the target variable as the central node and key driving variables as neighboring nodes, initial connection weights are generated based on response strength. An exponential decay factor is introduced to penalize the weights of long-lag paths, and weak connections are removed through pruning. This ultimately forms a sparse and physically meaningful graph prior matrix, serving as a guiding constraint for subsequent attention mechanisms. This step explicitly identifies the optimal response lag and dependency strength of each candidate variable on the target variable, thus providing structured prior knowledge consistent with industrial mechanisms for subsequent prediction models.

[0037] Furthermore, in some preferred embodiments of the present invention, a sampling reference time is predetermined; preprocessing includes: interpolating and aligning candidate variables at the sampling reference time; processing missing data by interpolation filling, interpolation of the mean under the same working condition, or direct removal according to a preset missing rate threshold; replacing outliers that deviate from the mean by more than a preset standard deviation multiple with the neighborhood median based on segmented statistics under working conditions; and normalizing candidate variables according to training set statistics.

[0038] Specifically, the purpose of the preprocessing stage is to construct a unified format input tensor from heterogeneous industrial data sources that can be directly consumed by subsequent deep learning models. In real-world industrial scenarios, raw SCADA data often suffers from problems such as inconsistent sampling, missing values, anomalous jumps, and significant differences in dimensions. Therefore, it must undergo a systematic data cleaning and standardization process.

[0039] First, to address the issue of inconsistent sampling frequencies, the sampling clock of key control variables (such as power and speed) is selected as the sampling reference time. Cubic spline interpolation is then performed on the remaining variables at the reference time to obtain a multivariate time series matrix with a unified sampling interval. For example, in the scenario of predicting the temperature of the main transformer winding of a wind turbine, the sampling clock of the active power of the turbine itself can be selected as the reference, and other telemetry variables can be time-aligned.

[0040] Secondly, for missing data, stratification is performed based on the missing rate: for variables with a missing rate of less than 5%, forward imputation and linear interpolation are combined to preserve local trends without introducing too much bias; for variables with a missing rate of more than 5% but less than 30%, mean imputation under similar operating conditions is used, that is, the mean is calculated from historical samples similar to the current operating conditions to fill the missing data, so as to maintain the consistency of operating conditions; variables with a missing rate of more than 30% are directly removed to avoid introducing too much unreliable information.

[0041] Furthermore, for abnormal jumps (such as instantaneous zero jumps or exceeding the measurement range by sensors), a 3σ criterion based on operating condition segments is used for identification: the continuous operating segment is divided into several operating condition clusters according to the load interval, and the mean and standard deviation are calculated within each operating condition cluster. Sample points exceeding 3σ are marked as abnormal and replaced with the neighborhood median. This approach fully considers the differences in data distribution under different operating conditions and avoids misjudgment of the global threshold.

[0042] Finally, to eliminate differences in the magnitude of the values ​​between variables (e.g., temperature is on the order of 10), 2 The current magnitude is 10 3 The pressure level is 10. 6 The impact of Pa on model gradient optimization is assessed by Z-score normalization of each variable based on training set statistics. For the observation value of the i-th variable at time t... Its normalized form is: ; in, and Let ε = 10 be the mean and standard deviation of the i-th variable on the training set. -6 This is a numerically stable term. The normalization parameter is fixed with the training set and directly called during deployment to avoid data leakage. After the above preprocessing, a multi-source observation vector is constructed at sampling time t: ; in, Let represent the observed value of the i-th variable at time t, and N be the total number of variables. The target variable to be predicted is selected as index q, and the target sequence is... Construct a sliding window history input matrix. Where L is the historical window length, dynamically determined based on the sampling frequency and the prediction target, with a recommended value of 64~512, and a default value of L=192; the prediction target is the future H-step sequence. .

[0043] This preferred embodiment effectively solves the common problems of asynchronous sampling, missing data, anomalies, and dimensional differences in multi-source industrial data through the above-mentioned hierarchical preprocessing strategy. It provides a high-quality and consistent data foundation for model training and online inference, and ensures the stability and reliability of subsequent dynamic response analysis and attention learning.

[0044] Furthermore, in some preferred embodiments of the present invention, a lag search set is pre-defined; key driving variables are selected from candidate variables through dynamic response lag analysis, including: shifting the candidate variable sequence of each lag in the lag search set and pairing it with the target variable sequence, calculating the absolute value of the Pearson correlation coefficient and the normalized mutual information based on histogram binning respectively; weighting and fusing the absolute value and normalized mutual information with preset weights to obtain the composite response intensity; selecting the lag that maximizes the composite response intensity as the optimal lag of the candidate variable, and selecting the largest composite response intensity as the response intensity of the candidate variable; using the mean of the response intensities of all candidate variables as an adaptive threshold, and retaining variables with response intensities greater than the adaptive threshold as key driving variables.

[0045] Specifically, dynamic response lag analysis is one of the core innovations of this invention. Its purpose is to identify the optimal response lag time of each candidate variable to the target variable a priori through statistical and information theory methods before entering the deep network, and to explicitly encode the physical transfer characteristics of the industrial system into the subsequent modeling process.

[0046] First, define the candidate lag set Ω = {1, 2, ..., τ_max}, where... ,default For each lag value τ in the set Ω, shift the entire candidate variable sequence forward by τ sampling steps and pair it with the current sequence of the target variable to construct a sample pair. .

[0047] For each pair of samples, two types of indicators are calculated simultaneously to account for both linear and nonlinear dependencies. The first type is the Pearson correlation coefficient. The first type reflects the linear correlation between variables and is sensitive to monotonic linear transfer (such as current-power relationships); the second type is normalized mutual information based on histogram binning. Mutual information estimation reflects arbitrary nonlinear dependencies between variables and is of great significance for nonlinear transfer characteristics commonly found in industrial systems, such as dead zones, saturation, and step jumps. The mutual information estimation employs a 16-interval binning strategy, with bin boundaries determined based on the 1% and 99th quantiles of the training set distribution to avoid interference from extreme values. The mutual information estimation formula is: ; Where B=16 is the number of boxes. For joint distribution frequency estimation, and For edge distribution frequency estimation. The final composite response intensity is obtained through weighted fusion: ; in, The weights are a fusion of linear and nonlinear values, which can be adjusted according to the specific scenario. The response intensity is averaged over the entire training set using a rolling window to ensure statistical stability.

[0048] Subsequently, for each candidate variable i, iterate through all lag values ​​in the set Ω, and select the lag that maximizes the response strength as the optimal lag for that variable. The corresponding maximum response intensity is denoted as . The physical meaning is: variable i is historically lagging. The values ​​taken in each sampling step, for the current target variable The impact is greatest; The physical meaning is: under optimal lag, variable i and the target variable The overall strength of the dependency between them.

[0049] To further eliminate weakly correlated or purely noisy variables, an adaptive threshold is introduced for variable selection: the mean of the response intensity of all candidate variables is used. As a screening threshold, only the set V of variables where Si > θ is retained. This adaptive threshold requires no manual parameter tuning and is robust to the number of variables and the scenario. This preferred embodiment achieves automatic discovery and lag time identification of key driving variables through response intensity calculation by fusing linear and nonlinear dual indicators and adaptive threshold screening, providing the model with input features and interpretable basis with clear physical meaning.

[0050] Furthermore, in some preferred embodiments of the present invention, the sparse graph prior matrix is ​​constructed based on the response intensity and optimal lag of the key driving variables, including: taking the target variable as the central node, normalizing the response intensity of each key driving variable to obtain the original weights; correcting the original weights based on the exponential decay factor associated with the optimal lag, using the corrected weights as the edge weights of the connections between variables, and removing connections with edge weights less than a preset pruning threshold to obtain the sparse graph prior matrix; wherein, during model training, the sparse graph prior matrix is ​​updated once every preset training period based on the statistics of the current training set.

[0051] Specifically, the purpose of constructing the sparse graph prior matrix is ​​to transform the statistical conclusions of dynamic response lag analysis into a sparse graph prior tensor that can be integrated with neural networks, so that the subsequent attention mechanism has the property of mechanism guidance.

[0052] First, construct a directed weighted graph with the target variable q as the central node and other retained variables as neighboring nodes. The response intensity... Based on this, the original weights are obtained through softmax normalization: ; Considering that longer delays lead to more complex physical transmission paths and lower information reliability (more interference factors may be mixed into historical information from distant sources), an exponential decay factor is introduced to correct the original weights: ; Among them, attenuation coefficient This increases the lag from 0 to The time weight decays to approximately 1 / e of its original value. This decay mechanism is essentially equivalent to imposing a priori confidence penalty on long-lag paths, reflecting the industry's understanding of information timeliness.

[0053] In addition, graph sparsification is performed: for all Edges are pruned to prevent meaningless weak connections from entering subsequent attention calculations. This ultimately yields a sparse graph prior matrix. Of these, only A non-zero edge is connected to the target variable node. In engineering implementation, A' is updated once every preset training cycle based on the current training set statistics during model training, and directly reused in the inference stage, resulting in extremely low computational overhead. This embodiment of the invention constructs a compact and physically meaningful graph prior matrix through response intensity normalization, exponential decay correction, and sparse pruning, providing a reliable prior anchor for the subsequent attention mechanism and effectively reducing dependence on data and the risk of spurious associations.

[0054] Step S104: Based on the historical sequence of key driving variables and the prior matrix of sparse graph, train the basic prediction model until the preset training completion conditions are met to obtain the prediction model; wherein, the basic prediction model includes: temporal encoding, graph prior constrained attention, and three-component decoupled prediction; the loss function of the basic prediction model is based on prediction uncertainty, mean square error, three-component orthogonalization loss and weight regularization joint constraint.

[0055] Specifically, this step is the model training phase. Its core is to use historical data to perform end-to-end training on the basic prediction model, which consists of a temporal coding module, a graph prior-constrained attention module, and a three-component decoupled prediction module. Simultaneously, it jointly learns the prediction uncertainty until the loss function converges or the preset training rounds are reached. The key driving variables—historical sequences and the sparse graph prior matrix—required for training are already prepared.

[0056] During training, the temporal coding module extracts temporal dynamic features from the historical observation sequences of each key driving variable, mapping the original sequences of varying lengths into fixed-dimensional high-order vector representations and stacking them to form a multivariate feature representation matrix, providing a unified representation space for subsequent modeling of relationships between variables. The graph prior-constrained attention module injects the sparse graph prior matrix as a soft constraint into the logits calculation of the self-attention mechanism in logarithmic form. This applies physical prior guidance on top of data-driven attention learning, enabling the model to maintain stable relationship learning even with high data noise or insufficient samples, avoiding spurious associations that contradict known mechanisms. The three-component decoupled prediction module explicitly decomposes the future H-step prediction into three physically meaningful components: a trend component, a dynamic response component, and a residual component. The trend component captures the low-frequency trend of the target variable through moving average extrapolation; the dynamic response component is obtained by mapping the attention output through a multilayer perceptron to model multivariate driving effects; and the residual component fits high-frequency random perturbations through a gated recurrent unit network. The sum of these three components constitutes the final prediction sequence. The loss function of the basic prediction model optimizes point prediction accuracy, uncertainty calibration quality, and the degree of decoupling among the three components by jointly constraining prediction uncertainty (negative log-likelihood loss), mean squared error, three-component orthogonalization loss, and weight regularization term. Through this step, the model can stably learn complex temporal dependencies between variables under the guidance of physical prior constraints, and significantly improve long-term prediction accuracy and robustness under multi-scale decoupled modeling.

[0057] Furthermore, in some preferred embodiments of the present invention, the temporal coding includes: extracting the historical observation sequence of each key driving variable, inputting it into a temporal coding module that adopts a causal convolution, dilated causal convolution, temporal convolution network or recurrent neural network structure, and generating a high-order representation vector; stacking all high-order representation vectors into a multivariate feature representation matrix, and allowing different key driving variables to share the same set of coding parameters.

[0058] Specifically, the purpose of temporal coding is to map the historical sequence of key driving variables, which has been selected and retained, into a feature representation of a unified dimension, which serves as the input to the subsequent graph prior constraint attention module.

[0059] Let the set of key driving variables be... It contains the number of variables. For any retained variable Extract its historical observation sequence of length L to form: ; Then, the historical sequence is input into the time-series coding module for feature extraction to obtain the higher-order representation vector corresponding to the variable: ; Where d represents the dimension of the encoded features. The temporal coding module adopts a temporal coding structure that satisfies temporal causality constraints to ensure that only historical information up to the current moment is used during the coding process, avoiding the leakage of future information. In specific implementations, the temporal coding module can employ structures such as causal convolution, dilated causal convolution, temporal convolutional networks, or recurrent neural networks. Alternatively, a Transformer encoder that utilizes only historical information can be used, achieving temporal causality constraints through causal masks. All of the above implementation methods can achieve the purpose of extracting sequence features, and users can choose according to their actual computing resources and accuracy requirements. Preferably, the same set of coding module parameters is shared for different variables to improve sample utilization efficiency, reduce the model parameter size, and enhance the model's ability to uniformly represent the local dynamic patterns of different industrial variables.

[0060] After encoding all retained variables, the feature vectors corresponding to each variable are stacked row by row to form a multivariate feature representation matrix: ; This matrix E serves as the input for the graph prior constraint attention calculation in the next step. The time-series encoding module extracts local fluctuation patterns, phased change trends, and dynamic evolution characteristics from the historical sequences of each key variable, providing rich temporal representations for subsequent modeling of interactions between variables. This embodiment efficiently compresses multidimensional time series into a compact feature matrix through parameter-sharing causal time-series encoding, preventing future information leakage and improving the model's generalization ability to different dynamic patterns of industrial variables.

[0061] Furthermore, in some preferred embodiments of the present invention, graph prior-constrained attention includes: mapping the multivariate feature representation matrix to a query matrix, a key matrix, and a value matrix through a trainable projection; injecting the sparse graph prior matrix into the attention in logarithmic form, applying an adaptive sparsity strategy based on the number of variables, and renormalizing the retained attention connections; using the renormalized attention weights to perform a weighted summation of the value matrix, stacking multiple layers and introducing residual connections and layer normalization between layers to obtain the attention output.

[0062] Specifically, graph prior-constrained attention is another core innovation of this invention. The sparse graph prior matrix A' constructed in step S102 is used as a soft constraint to inject into the standard attention mechanism, so as to organically integrate data-driven learning with mechanism prior.

[0063] First, construct the Query, Key, and Value matrix according to standard practice: , , ,in, This is a trainable projection matrix. The key modification in this embodiment of the invention lies in the construction method of the attention logits: ; Where λ=0.3 is the graph prior fusion weight, and ε=10 -8 To prevent the logarithm from being negative infinity, logarithmic injection is used instead of linear superposition because the logits space before softmax is isomorphic to the log probability space. Logarithmic injection is equivalent to performing a Bayesian combination of prior and likelihood in the probability space, which is more theoretically reasonable. At the same time, the logarithmic term strongly suppresses zero-weight edges (close to -∞), naturally achieving sparsity.

[0064] Furthermore, to improve computational efficiency and interpretability in scenarios with large-scale variables, an adaptive sparsity strategy based on the number of variables is introduced: for each row of the attention matrix, only the row with the largest weight is retained. The remaining terms are set to zero and renormalized. This strategy results in k≈5 when N=20 and k≈7 when N=100, significantly reducing unnecessary weak connections compared to complete attention while preserving the main information channels.

[0065] The graph prior-constrained attention module consists of two stacked layers, with residual connections and LayerNorm used between layers. The core innovation of this mechanism lies in the fact that, when training data is sufficient and statistically reliable, data-driven terms... The graph prior is dominant, and the model can learn complex relationships beyond the prior; when the data is noisy or the sample is insufficient, the graph prior is dominant. It acts as an anchor, constraining attention weights within a physically reasonable range to avoid spurious associations; the weight λ controls the balance between the two, and can be used as a hyperparameter for tuning during training, or designed as a learnable parameter for adaptive adjustment. This invention, through prior injection and adaptive sparsity, achieves an organic combination of data-driven and mechanism-guided approaches, significantly enhancing the robustness and interpretability of attention learning.

[0066] Furthermore, in some preferred embodiments of the present invention, the three-component decoupled prediction includes: performing a moving average on the historical sequence of the target variable, and linearly extrapolating the trend sequence obtained after the moving average to generate a trend component; flattening the attention output of the graph prior constraint attention output and inputting it into a multilayer perceptron to map and generate a dynamic response component; inputting the historical residual sequence into a gated recurrent unit network and linearly mapping it to generate a residual component; wherein, the historical residual sequence is obtained by subtracting the estimated values ​​of the historical parts corresponding to the trend component and the dynamic response component from the historical sequence of the target variable; and summing the trend component, the dynamic response component, and the residual component to form a prediction sequence.

[0067] Specifically, the three-component decoupled prediction explicitly decomposes the future H-step prediction into three mutually independent components with clear physical meanings, each handled by a dedicated submodule, thereby reducing the fitting burden of a single network and alleviating the accumulation of long-term time-domain errors.

[0068] The overall forecast formula is: ; (1) Trend component T: Responsible for characterizing the low-frequency trend of the target variable (such as load ramp-up and temperature drift on an hourly scale), using a lightweight scheme of moving average plus least squares extrapolation. Specifically, for the target variable sequence within the historical window... Application window length The moving average operator is used to obtain the trend sequence. The slope is obtained by fitting a first-order polynomial to its nearest w points. Extrapolation yields the future H-step trend: The formula of this scheme is simple, the extrapolation is stable, and it will not be contaminated by high-frequency noise.

[0069] (2) Dynamic response component D: Responsible for characterizing the future changes of the target variable driven by other variables, obtained from the output of the prior constraint attention module in step S104 via a linear mapping head. In specific implementation, the attention output tensor is... After flattening, an H-dimensional vector is obtained through two layers of MLP: the first layer has a dimension of N·d→256, GELU activation, and Dropout=0.1; the second layer has a dimension of 256→H. This submodule is responsible for the main physical causal modeling and is the core source of prediction accuracy.

[0070] (3) Residual component E: Responsible for modeling the high-frequency random disturbances remaining after removing the trend and dynamic response. Specifically, the residual sequence is defined. Extract the most recent The residual is used as input and fed into a two-layer GRU network (64 hidden dimensions), finally mapped to an H-dimensional output through a linear layer. Although the residual component accounts for a small proportion of the total prediction magnitude, it is crucial for capturing short-term fluctuations and can significantly reduce the mean squared error of the prediction. This embodiment of the invention uses a physically-meaning-oriented three-component decoupling design to divide and conquer the originally coupled multi-scale modeling task, effectively reducing the ineffective consumption of model capacity and alleviating the problem of error accumulation in long-term prediction.

[0071] Furthermore, in some preferred embodiments of the present invention, the loss function of the basic prediction model is based on the joint constraints of prediction uncertainty, mean square error, three-component orthogonalization loss, and weight regularization, including: setting an independent branch to output variance correlation in the prediction head of the dynamic response component; calculating the negative log-likelihood loss based on the variance correlation; and weighting and summing the negative log-likelihood loss, mean square error loss, three-component orthogonalization loss, and weight regularization term as the joint loss function of the basic prediction model.

[0072] Specifically, in the loss function design and model training stages, we not only focus on the accuracy of point prediction, but also quantify risk by outputting prediction uncertainty.

[0073] In the prediction head of the dynamic response component, an independent branch is set: the output dimension of the last layer of the original response branch is expanded to 2H, and the first H dimensions are the prediction mean. The H-dimensional dimension is the predicted log-variance. The reason for using logarithmic variance instead of direct variance is that variance is naturally positive, and direct regression may result in negative values, leading to training divergence; logarithmic variance can be freely optimized over the entire real domain, and then restored to positive variance through exponential transformation, significantly improving numerical stability. The mean head and variance head share the underlying attention features, but are mapped separately through two independent linear projections to avoid mutual interference between tasks.

[0074] During the training phase, negative log-likelihood loss (NLL) under the Gaussian distribution assumption is used. ; The NLL loss exhibits good adaptive properties: when the prediction error is large at a certain step, the model can be increased... This reduces the loss contribution of the first term, thus enabling self-reporting of high uncertainty intervals; however, increasing variance will simultaneously increase the second term. The cost of this is such that the model will not expand the variance indefinitely, but will achieve an optimal balance between the two.

[0075] The three-component orthogonalization loss is used to avoid mutual interference between components. It is calculated as the sum of the absolute values ​​of the cosine similarity between each pair of the predicted sequences of the three components: ; The final total loss function is a weighted combination of multiple objectives: ; The first term is the uncertainty perception loss, providing a variance signal; the second term is the mean squared error, accelerating point prediction convergence and avoiding variance explosion in the early stages of training; the third term is the three-component orthogonalization loss; and the fourth term is the L2 regularization of the model weights. The weights were optimized experimentally to γ1=0.5, γ2=0.1, and γ3=10. -4 The training optimizer uses AdamW with an initial learning rate of 1×10⁻⁶. -4 The algorithm employs cosine annealing scheduling, a batch size of 64, and 100 epochs, with early stopping (patience=10) to prevent overfitting. When the joint loss function converges or reaches the preset number of iterations, the preset training completion condition is met, and a well-trained prediction model is obtained. This embodiment of the invention achieves simultaneous improvement in prediction accuracy and risk quantification capability by jointly optimizing point prediction error, uncertainty calibration, component orthogonality, and weight regularization. The training process is stable and less prone to overfitting.

[0076] Step S106: The real-time running data is preprocessed and then input into the trained prediction model to obtain the prediction sequence and the confidence level of each step in the prediction sequence. Online correction is performed when the confidence level is less than the preset confidence level threshold. Rolling calibration is performed based on the recent prediction performance after the prediction model is deployed.

[0077] Specifically, this step is the model deployment and online prediction phase. Its purpose is to use the trained prediction model to perform long-term time-domain predictions on real-time data and adaptively perform corrections based on prediction confidence. Simultaneously, a rolling calibration mechanism ensures the model's performance stability during long-term operation. The prediction model receives real-time data that has undergone the same preprocessing as in the training phase and outputs a prediction sequence for the next H steps, along with the confidence level for each step. The confidence level is obtained by exponentially transforming the prediction variance output synchronously by the model, reflecting the model's grasp of the predicted value at each step. When the confidence level at a certain step falls below a preset threshold, it indicates that the main model's prediction reliability at that moment is insufficient. This triggers a weighted fusion correction using a ridge regression short-term model trained based on recent real observations and generated prediction values. The short-term model's ability to quickly track local changes compensates for the main model's inability to adapt to sudden changes in operating conditions. Furthermore, the system automatically evaluates the quality of recent predictions at preset intervals, detecting systematic biases in the prediction distribution using two indicators: continuous hierarchical probability scoring and experience coverage. When the bias exceeds the limit, the confidence threshold is dynamically adjusted to maintain the effectiveness of the correction mechanism, thereby automatically adapting to long-term distribution drift caused by equipment aging, seasonal changes, and adjustments in operating modes. Through this step, the prediction model can continuously output high-precision predictions in real-world operating environments and maintain robust adaptability to uncertainties.

[0078] Furthermore, in some preferred embodiments of the present invention, online calibration includes: The confidence level of each step in the prediction sequence is calculated based on the variance correlation quantity output by the prediction model, so that the confidence level decreases monotonically as the uncertainty increases. When the confidence level is less than the preset confidence level threshold, the regularized linear model trained based on recent measured values ​​and the previously generated prediction values ​​in the prediction sequence is weighted and fused with the prediction mean of the prediction sequence to obtain the corrected prediction value.

[0079] Specifically, the online correction mechanism includes three stages: confidence calculation, trigger determination, and fusion of long-term and short-term models.

[0080] After the model is deployed, for any prediction of the next H steps at any current time t, each step is associated with a prediction log-variance. The standard deviation is obtained through exponential transformation. The confidence level of the prediction at step h is defined as: ; This definition monotonically maps the standard deviation from the [0,+∞) interval to the (0,1] confidence interval: the confidence level approaches 1 (absolutely reliable) when the standard deviation approaches 0, and the confidence level approaches 0 (completely unreliable) when the standard deviation approaches infinity. This definition is simple in form, engineer-friendly, and decoupled from the normalized target variable's dimensions, facilitating unified comparisons across variables and scenarios.

[0081] Set a base threshold It can be adjusted between 0.5 and 0.7 depending on the actual scenario. When If the main model is deemed unreliable at the prediction point, an online correction mechanism is triggered; otherwise, the predicted mean output by the main model is used directly. .

[0082] The core of the correction mechanism is to construct a short-term model as a supplement to the main model. The short-term model employs ridge regression based on observed and predicted values: using the most recent p-step measured values... Compared with the predicted first h-1 steps The input features (p+h-1 dimensions in total, default p=12) are used to... For output, train a linear mapping offline. The regularization coefficient is 0.01. The corrected predicted value is defined as: ; This correction mechanism is equivalent to a Bayesian weighted fusion that adaptively balances the long-term complex model and the short-term simple model based on the prediction uncertainty: when the confidence of the main model is high (such as in a stationary condition). The main model dominates the prediction; when the confidence of the main model is low (such as during a change in operating conditions or a sudden disturbance). The short-term model dominates the prediction, thus avoiding large errors in scenarios where the main model is not well-suited. As an extended implementation, the short-term model can also use support vector regression or a lightweight fully connected network, as long as it meets the conditions of fast training and low online inference latency. This embodiment of the invention effectively combines the long-range modeling capability of deep models with the local fast response capability of linear models through confidence-driven adaptive fusion, significantly reducing the error accumulation at the final step of long-term prediction and the instantaneous prediction error in scenarios with changing operating conditions.

[0083] Furthermore, rolling calibration includes: after the prediction model is deployed, at preset time intervals, calculating the calibration score and empirical coverage of the prediction distribution based on recent measured values, historical predicted values ​​and variance correlation quantities; when the empirical coverage deviates from the preset ideal coverage by more than the preset allowable deviation, adjusting the confidence base threshold within the protection interval by a preset step size.

[0084] Specifically, the rolling calibration mechanism is designed to address long-term distribution drift issues such as equipment aging, seasonal changes, and adjustments to operating modes, and to automatically maintain model performance.

[0085] During online operation, the system automatically performs a calibration every preset interval (default 24 hours): calibrating the most recent measured values. Corresponding historical forecast values Prediction variance Pairing. The window length W is determined based on the sampling frequency. For example, in a wind turbine SCADA system with a sampling interval of 1 minute, W=1440 corresponds to 24 hours.

[0086] Two key calibration metrics are calculated. The first metric is the Continuous Graded Probability Score (CRPS), which measures the overall fit between the predicted distribution and the actual observations: ; in, Let be the cumulative distribution function corresponding to the i-th prediction, which has an analytical solution under the Gaussian assumption, facilitating rapid calculation. The second indicator is the empirical coverage rate, i.e., the measured value falling within the 90% confidence interval of the prediction. The ideal sample proportion is 0.9.

[0087] If the CRPS is significantly high or the empirical coverage deviates from 0.9 by more than ±5%, then a systematic bias in variance prediction is identified. If the coverage is too low (e.g., <0.85), it indicates that the variance is underestimated, and the baseline confidence threshold needs to be increased. To increase the frequency of calibration triggering; excessively high coverage (e.g., >0.95) indicates that the variance is overestimated, and C0 needs to be lowered to reduce unnecessary calibration. The default adjustment step size is 0.05, and the range of a single adjustment is limited to 0.4~0.8 to prevent instability. This mechanism enables the present invention to automatically address long-term distribution drift issues such as equipment aging (e.g., temperature baseline drift caused by bearing wear), seasonal changes (e.g., differences in ambient temperature between summer and winter), and adjustments to operating modes (e.g., changes in peak shaving strategies), avoiding the performance degradation of traditional models over time after deployment and significantly extending the effective service life of the model. In engineering implementation, the rolling calibration module is executed asynchronously as an independent process, without affecting the real-time performance of the main prediction link. This preferred embodiment achieves long-term adaptation of the prediction model in non-stationary industrial environments through dynamic threshold adjustment based on CRPS and coverage feedback, ensuring the reliability of the prediction service throughout its service life.

[0088] The long-term prediction method for industrial equipment state variables provided by this invention explicitly characterizes the transmission characteristics of industrial mechanisms, improving the interpretability of predictions. Through Dynamic Response Lag Modeling (DRLM), a systematic approach is adopted to search for the optimal lag of each variable using a dual-index approach of "linear correlation + nonlinear mutual information." This enables the model to accurately identify real physical phenomena such as "feedwater flow rate leading drum water level by 30 seconds" and "paddle pitch angle leading main bearing temperature by 2 minutes." The optimal lag and response intensity of each variable can be used as interpretive outputs of the model, facilitating understanding of the model's decision-making basis by maintenance engineers and overcoming the "black box" deficiency of traditional deep models.

[0089] The long-term prediction method for industrial equipment state variables provided by this invention exhibits more stable and robust attention learning and reduced noise sensitivity. By injecting the sparse graph prior into the attention logits in logarithmic form and combining it with Top-k sparsity, the model no longer degenerates into pseudo-associations in small sample and strong noise scenarios, and the attention weights remain stable under operating condition switching conditions. Experiments on a wind turbine measured dataset show that the KL divergence of the attention weights between operating conditions is reduced from 0.42 in the baseline method to 0.11, improving stability by approximately 74%.

[0090] The long-term prediction method for industrial equipment state variables provided by this invention significantly improves the long-term prediction accuracy. By decoupling the trend-response-residual components and orthogonalizing the loss, the originally coupled multi-scale modeling task is divided and conquered. In the experiment of predicting the water level of a thermal power plant steam drum (H=96), this method reduces the MSE by approximately 23.6% compared to Informer, approximately 17.9% compared to Autoformer, and approximately 12.4% compared to iTransformer; in the prediction of wind power main bearing temperature, the MAE of the 96-step prediction is reduced by approximately 15.5% compared to the optimal baseline.

[0091] The long-term prediction method for industrial equipment state variables provided by this invention effectively suppresses error accumulation. The confidence-driven correction mechanism reduces the error growth rate at the final step (step H) of the long-term prediction from 60%–80% to below 25%, significantly improving the usability of long-term predictions. The maximum instantaneous error at the moment of operating condition switching is reduced from 21.4 mm in the original method to 9.7 mm (steam drum water level case), a reduction of 54.7%.

[0092] The long-time domain prediction method for industrial equipment state variables provided by this invention has the ability to quantify uncertainty, supporting risk decision-making. The prediction variance is output synchronously. It can directly provide risk quantification basis for downstream operation decisions (such as dynamic adjustment of safety thresholds, start-up and shutdown of standby units, and preparation of maintenance plans), and is easy to integrate with existing industrial SIS / MIS systems to form a complete closed loop of "prediction + confidence + decision".

[0093] The long-term time-domain prediction method for industrial equipment state variables provided by this invention supports adaptive long-term distribution drift. The rolling calibration mechanism enables this invention to address long-term distribution drift issues such as equipment aging, seasonal changes, and adjustments to operating modes, avoiding the performance degradation of traditional models over time after deployment and extending the effective service life of the model.

[0094] For example, consider its application in long-term time-domain prediction of main transformer winding temperature.

[0095] I. Experimental subjects and data sources.

[0096] In this embodiment, the predicted target state quantity is the main transformer winding temperature. Candidate input variables include: active power, reactive power, power factor, frequency, A-phase current, B-phase current, C-phase current, high-voltage reactive power, A-phase current, B-phase current, and C-phase current.

[0097] The experimental data spanned from 2:39 AM on July 27, 2025 to 1:33 AM on August 29, 2025, with a sampling interval of 1 minute. The collected data from each measurement point were aligned along a unified time axis to form a multivariate time series sample.

[0098] The data was divided as follows: the first 60% of the samples were used as the training set; the middle 20% of the samples were used as the validation set; and the last 20% of the samples were used as the test set.

[0099] To verify the stability and applicability of this invention under different prediction lengths, the following two long-term prediction tasks were set up: This means predicting the main transformer winding temperature sequence for the next 60 minutes. This means predicting the main transformer winding temperature sequence for the next 90 minutes.

[0100] Let the target variable be the temperature of the main transformer winding at time t. The observed values ​​are denoted as There are a total of candidate driving variables The, the The variables at time... The observed values ​​are denoted as Then, within a given history window length... Under these conditions, the model input can be represented as: ; Predicting the target is the future. Main transformer winding temperature sequence: ; Multivariate time series construction and normalization.

[0101] First, the aforementioned multi-source SCADA telemetry variables are aligned according to a unified timestamp. For a small number of missing points, interpolation is used to fill in the gaps, thereby constructing a multivariate time series sample consisting of the target variable and candidate driving variables.

[0102] Subsequently, the variables were standardized based on the training set statistics to eliminate the impact of differences in the units of measurement of different variables on the model training and prediction results. For any variable... Its standardized form is: ; in, and These represent the mean and standard deviation of the variable on the training set, respectively. Both the validation and test sets utilize the statistics from the training set during processing to ensure consistency across different stages and prevent data leakage.

[0103] Based on the above preprocessing results, supervised learning samples are constructed. For any given time... Take the length as Using historical windows as input, to extract the future The main transformer winding temperature sequence at each time point is used as the prediction target, i.e.: ; in, ; Dynamic response hysteresis analysis.

[0104] Because the temperature of the main transformer winding has significant thermal inertia, its change usually lags behind changes in driving variables such as load and current. Therefore, before formal modeling, the dynamic response lag relationship between each candidate variable and the target variable is analyzed.

[0105] For each candidate variable Set the lag search range as At each lag step Below, the Pearson correlation coefficient and normalized mutual information between the Pearson correlation coefficient and the target variable, the main transformer winding temperature, are calculated respectively, and the results are as follows: ; ; in, Used to measure linear correlation. This is used to measure nonlinear dependencies. To balance both aspects, a composite response strength index is constructed: ; in, This is the weighting coefficient. In one implementation, it can be taken as... That is, linear correlation and nonlinear dependence are given the same weight.

[0106] For each candidate variable, the maximum composite response strength within the lag search range is taken as the effective response strength of that variable: ; The lag step size corresponding to the attainment of the maximum value is recorded as the optimal response lag: ; according to The size of all candidate variables is used to sort them, and a set of key driving variables is selected by combining a preset threshold or a Top-K strategy. .

[0107] After analysis, the following key driving variables were automatically selected in this embodiment: high-voltage C-phase current, high-voltage B-phase current, high-voltage A-phase current, body frequency, body active power, and body A-phase current; among which: the optimal lag of the high-voltage three-phase current is about 19 to 22 minutes; the optimal lag of the body active power is about 15 minutes; and the optimal lag of the body frequency and body A-phase current is about 1 minute.

[0108] The results show that the method of the present invention can automatically identify the objective mechanism relationship of "load change occurs first, and the main transformer winding temperature responds subsequently", thereby providing a physically meaningful structural prior for the subsequent prediction model.

[0109] Construction of predictive inputs based on key variable screening.

[0110] After obtaining the key driving variables and their optimal lag information, all candidate variables are no longer input into the prediction model indiscriminately. Instead, only the key variables selected through dynamic response analysis are retained to reduce the interference of redundant and noisy variables on the prediction results.

[0111] Suppose that the set of key variables obtained after filtering is: ,in, Indicates the number of key variables.

[0112] For any key variable Extract the observation sequence within its historical window to form: ; Furthermore, in a preferred embodiment, optimal response hysteresis may also be incorporated. Aligning or weighting the input window can enhance the model's ability to express key dynamic delay relationships. For example, weights based on hysteresis strength can be applied to different variables. It can be obtained by normalizing the effective response intensity of the variable: ; By retaining only key variables and introducing lagged priors, the spurious effects of weakly correlated variables can be reduced while preserving the main driving information, thereby improving the model's ability to characterize the temperature variation of the main transformer winding.

[0113] Key variable sequence encoding.

[0114] To further extract dynamic patterns from the historical sequences of key driving variables, the historical sequences of each retained variable are input into the time-series coding module and mapped to a high-order feature representation of a unified dimension.

[0115] For any key variable The encoding result is represented as: ;in, Indicates the sequence encoding module. For feature dimensions.

[0116] The sequence encoding module preferably employs a temporal encoding structure that satisfies temporal causality constraints, ensuring that only current and historical information is used during the encoding process without introducing future information. It can be implemented using causal convolution, dilated causal convolution, temporal convolutional networks, recurrent neural networks, or other structures suitable for time series modeling.

[0117] After encoding all key variables, the feature vectors of each variable are stacked to form a multivariate feature representation matrix: ; matrix It is used to characterize the dynamic state of each key variable within the current historical window and serves as input for subsequent variable relationship modeling and prediction decoding.

[0118] Modeling of variable relationships based on lag priors.

[0119] Since different key variables have varying degrees of influence and coupling relationships on the temperature of the main transformer winding, it is necessary to further model the relationship structure between variables after obtaining the variable encoding representation.

[0120] In this embodiment, a priori matrix of variable relationships is constructed based on the dynamic response analysis results obtained in step 2. For any two key variables and The strength of their prior association can be determined based on the strength of their responses to the target variable, their lag proximity, or their statistical correlation. One possible representation is: ; in, This indicates the prior constructor function.

[0121] In a preferred embodiment, it can be set as follows: ;in, This parameter controls the sensitivity to lag differences. Therefore, when the optimal lags of two variables to the target variable are similar and their response strengths are high, their prior association weights are greater.

[0122] Furthermore, this prior matrix can be used to constrain the learning process of relationships between variables, making the model more inclined to retain structural connections that conform to dynamic response laws while focusing on data-driven relationships, thereby enhancing the interpretability and robustness of the model.

[0123] Long-term prediction output.

[0124] After obtaining the high-order feature representations and relational structures of the key variables, they are input into the prediction module to output the future. Predicted temperature values ​​for the main transformer winding: ;in, This represents the prediction mapping function, with the variable feature matrix as input. and their prior relationships The output is the future prediction sequence: In engineering implementation, the prediction module can employ fully connected mapping, sequence decoders, multi-step parallel output heads, or autoregressive decoding structures. To improve the stability of long-term prediction, this embodiment preferably adopts a multi-step direct output method to generate future predictions in one step. Step-by-step prediction results are used to mitigate the problem of gradual error accumulation.

[0125] Model training and evaluation metrics.

[0126] The model parameters are trained using the training set, hyperparameters are selected and early stopping is controlled using the validation set, and the model performance is finally evaluated on the test set.

[0127] The training objective uses the mean squared error loss function between the predicted and true values: ; During the testing phase, the root mean square error (RMSE), mean absolute error (MAE), and terminal prediction error were used to evaluate the model performance. Specifically: ; ; in, This represents the total number of predicted points on the test set. The terminal RMSE measures the error level at the last moment of the prediction interval to reflect the model's stability in long-term predictions.

[0128] Furthermore, the embodiments of the present invention provide the following comparison methods: 1. Continuous prediction; 2. Moving average extrapolation; 3. XGBoost; 4. LSTM; 5. Informer; 6. Autoformer; 7. The method provided in the embodiments of the present invention. Among these, Informer and Autoformer employ lightweight implementations adapted to the current sample size to meet the experimental requirements of reproducibility and comparability in engineering scenarios. It should be noted that continuous prediction and moving average extrapolation belong to traditional statistical / empirical prediction baselines, XGBoost represents a typical machine learning method, LSTM represents a classic recurrent neural network method, and Informer and Autoformer represent deep learning methods in the field of long sequence prediction. Therefore, the above comparisons have good hierarchy and representativeness.

[0129] See Figure 2 The diagram shown is a comparison of the overall RMSE of multiple prediction methods for a prediction time domain of 60 hours, provided by an embodiment of the present invention. Figure 3 The diagram shown in this embodiment of the invention illustrates the overall MAE comparison results of multiple prediction methods when the prediction time domain is 60. It can be seen that, under the condition H=60, the method of the present invention outperforms all the comparison methods in terms of RMSE, MAE, and terminal RMSE. (See also...) Figure 4 The diagram shown is a comparison of the overall RMSE of multiple prediction methods for a 90-hour prediction time domain, provided by an embodiment of the present invention. Figure 5 The schematic diagram shown in this embodiment of the invention provides a comparison of the overall MAE results of various prediction methods when the prediction time domain is 90. It can be seen that, under the condition H=90, the method of the present invention still maintains the best overall RMSE and overall MAE indicators. Although LSTM has a slightly lower single-step error at the end than the method of the present invention under the condition H=90, the method of the present invention still has the best overall performance in terms of the overall error performance over the entire prediction interval. This indicates that the advantage of the method of the present invention is mainly reflected in the overall long-term prediction stability and comprehensive error control capability, and not just in the local error of a single prediction step. Furthermore, the overall MAE indicator shows that the method of the present invention achieves the lowest value under both prediction lengths, indicating that the method also has a stable advantage in terms of mean absolute error.

[0130] The joint analysis of the key variable screening results and prediction performance reveals that load-related variables such as high-voltage three-phase current and active power have a significant driving effect on the temperature of the main transformer winding. The optimal response lag of these variables is mainly distributed in the range of 15 to 22 minutes, which is consistent with the delayed temperature rise mechanism caused by the thermal inertia of the main transformer. After screening key variables based on this type of lag prior, the noise disturbance caused by low-correlation variables can be effectively reduced. Under long-term conditions, the model can more stably capture the evolution relationship of "load change - thermal response - temperature change", so the overall prediction effect is better than that of the general model.

[0131] In other words, the method of this invention does not rely solely on black-box fitting, but first extracts the physically meaningful variable-target coupling relationship through dynamic response analysis, and then uses this relationship for subsequent prediction modeling. Therefore, it is more adaptable to long-term time-domain prediction scenarios of industrial equipment state variables.

[0132] The long-time domain prediction method for industrial equipment state variables provided by this invention can automatically identify key driving variables and their optimal lag relationships, thereby improving model interpretability. This invention not only outputs the prediction results of the main transformer winding temperature, but also identifies the time delay characteristics of the influence of different driving variables on the target quantity, making the model results more consistent with the thermal inertia of the main transformer and the load transfer mechanism, thus significantly enhancing the interpretability of the prediction process.

[0133] The long-term prediction method for industrial equipment state variables provided by this invention can eliminate invalid variables and redundant noise before modeling, reduce spurious correlation interference, and through dynamic response lag analysis and key variable screening, this invention can eliminate a large number of weakly correlated or invalid variables before entering the prediction model, reduce the interference of spurious related information, noise information and redundant information on the prediction model, and improve the effectiveness of input features and the robustness of the model.

[0134] The long-term prediction method for industrial equipment state variables provided by this invention has significant comprehensive performance advantages in long-term prediction tasks. In two long-term tasks with H=60 and H=90, the method of this invention outperforms the comparative methods such as continuous prediction, moving average extrapolation, XGBoost, LSTM, Informer and Autoformer in terms of overall RMSE and overall MAE, indicating that it has better prediction accuracy and stability under long-term conditions.

[0135] The long-time domain prediction method for industrial equipment state variables provided by this invention is more suitable for industrial equipment state variable prediction scenarios with dynamic response lag and complex operating condition disturbances. For industrial state variables such as main transformer winding temperature, which have thermal inertia, lag transmission and multivariable coupling characteristics, the method of this invention can extract key driving relationships more effectively and enhance the model's adaptability to complex operating condition changes. Therefore, it is more suitable for long-time domain prediction applications of actual industrial equipment state variables.

[0136] The long-term time-domain prediction method for industrial equipment state variables provided by this invention combines engineering feasibility with methodological scalability. This method can be implemented using only conventional SCADA telemetry data from industrial sites, without requiring additional expensive sensors or complex experimental conditions, making it easy to deploy and apply in existing booster stations, substations, and other industrial monitoring scenarios. Furthermore, this method can also be extended to predicting oil temperature, core temperature, equipment load status, and other industrial state variables exhibiting hysteresis dynamic characteristics.

[0137] In summary, for the task of long-term prediction of main transformer winding temperature, this invention introduces an overall technical route of dynamic response lag analysis, key variable screening, and sequence modeling prediction. This approach can effectively identify key driving factors and their delay effects, reduce interference from redundant inputs, and improve the overall accuracy and stability of long-term prediction.

[0138] Therefore, the method of the present invention has good application value and promotion prospects in the prediction of state variables of industrial equipment such as main transformers with significant thermal inertia and multivariable coupling characteristics.

[0139] This invention provides a long-term prediction method for state variables of industrial equipment. It explicitly identifies the optimal lag and response strength of each variable to the target variable through dynamic response lag analysis, accurately characterizing the physical transfer characteristics of the industrial system and making the model interpretable. By injecting a sparse graph prior matrix constructed based on response strength and optimal lag as a soft constraint into the attention mechanism, attention learning becomes more stable and robust, reducing noise sensitivity. Three-component decoupled prediction separates and conquers the trend, dynamic response, and residuals, reducing the modeling burden and improving long-term prediction accuracy. Joint learning predicts uncertainty and outputs confidence scores, with an online correction mechanism effectively suppressing error accumulation and improving long-term prediction reliability. Rolling calibration automatically adapts to equipment state distribution drift, extending the effective service life of the model.

[0140] Based on the above embodiments, this invention provides a long-term time-domain prediction device for industrial equipment state variables, see [link to relevant documentation]. Figure 6 The diagram shown is a structural schematic of a long-time domain prediction device for industrial equipment state variables provided in an embodiment of the present invention. The device includes: The data processing module 310 is used to acquire and preprocess historical multi-source time series data of industrial equipment, select key driving variables from candidate variables through dynamic response lag analysis, and construct a sparse graph prior matrix based on the response intensity and optimal lag of the key driving variables. The model training module 320 is used to train the basic prediction model based on the historical sequence of key driving variables and the sparse graph prior matrix until the preset training completion conditions are met to obtain the prediction model. The basic prediction model includes: temporal encoding, graph prior constrained attention, and three-component decoupled prediction. The loss function of the basic prediction model is based on prediction uncertainty, mean square error, three-component orthogonalization loss, and weight regularization joint constraint. The data prediction module 330 is used to input the preprocessed real-time running data into the trained prediction model to obtain the prediction sequence and the confidence level of each step in the prediction sequence, and to perform online correction when the confidence level is less than the preset confidence level threshold; wherein, after the prediction model is deployed, rolling calibration is performed based on the recent prediction performance.

[0141] Furthermore, in some preferred embodiments of the present invention, a sampling reference time is predetermined; the data processing module 310 is used to interpolate and align the candidate variables at the sampling reference time, and process the missing data by interpolation filling, mean interpolation under the same working condition, or direct removal according to a preset missing rate threshold; based on the segmented statistics of working conditions, outliers that deviate from the mean by more than a preset standard deviation are replaced by the neighborhood median; and the candidate variables are normalized according to the statistics of the training set.

[0142] Furthermore, in some preferred embodiments of the present invention, a lag search set is pre-defined; the data processing module 310 is used to shift the candidate variable sequence of each lag in the lag search set and pair it with the target variable sequence, calculate the absolute value of the Pearson correlation coefficient and the normalized mutual information based on histogram binning respectively; the absolute value and normalized mutual information are weighted and fused by preset weights to obtain the composite response intensity; the lag that maximizes the composite response intensity is taken as the optimal lag of the candidate variable, and the largest composite response intensity is taken as the response intensity of the candidate variable; the mean of the response intensity of all candidate variables is used as the adaptive threshold, and variables with response intensity greater than the adaptive threshold are retained as key driving variables.

[0143] Furthermore, in some preferred embodiments of the present invention, the data processing module 310 is used to obtain the original weights by normalizing the response intensity of each key driving variable with the target variable as the central node; to correct the original weights based on the exponential decay factor related to the optimal lag; to use the corrected weights as the edge weights of the connections between variables; and to remove connections with edge weights less than a preset pruning threshold to obtain the sparse graph prior matrix; wherein, during the model training process, the sparse graph prior matrix is ​​updated once every preset training cycle based on the current training set statistics.

[0144] Furthermore, in some preferred embodiments of the present invention, the model training module 320 is used to extract the historical observation sequence of each key driving variable, input it into a temporal coding module that adopts a causal convolution, dilated causal convolution, temporal convolution network or recurrent neural network structure, generate a high-order representation vector; stack all high-order representation vectors into a multivariate feature representation matrix, and allow different key driving variables to share the same set of coding parameters.

[0145] Furthermore, in some preferred embodiments of the present invention, the model training module 320 is used to map the multivariate feature representation matrix into a query matrix, a key matrix, and a value matrix through a trainable projection; inject the sparse graph prior matrix into the attention in logarithmic form, apply an adaptive sparsity strategy based on the number of variables, and renormalize the retained attention connections; use the renormalized attention weights to perform a weighted summation of the value matrix, stack multiple layers, and introduce residual connections and layer normalization between layers to obtain the attention output.

[0146] Furthermore, in some preferred embodiments of the present invention, the model training module 320 is used to perform a moving average on the historical sequence of the target variable, and to perform linear extrapolation on the trend sequence obtained after the moving average to generate a trend component; the attention output of the graph prior constraint attention output is flattened and input into a multilayer perceptron to map and generate a dynamic response component; the historical residual sequence is input into a gated recurrent unit network and linearly mapped to generate a residual component; wherein, the historical residual sequence is obtained by subtracting the estimated values ​​of the historical parts corresponding to the trend component and the dynamic response component from the historical sequence of the target variable; the trend component, the dynamic response component and the residual component are summed to form a prediction sequence.

[0147] Furthermore, in some preferred embodiments of the present invention, the model training module 320 is used to set an independent branch output variance correlation quantity in the prediction head of the dynamic response component; calculate the negative log-likelihood loss based on the variance correlation quantity, and weight the negative log-likelihood loss, mean squared error loss, three-component orthogonalization loss and weight regularization term by summing them as the joint loss function of the basic prediction model.

[0148] Furthermore, in some preferred embodiments of the present invention, the data prediction module 330 is used to calculate the confidence level of each step in the prediction sequence based on the variance correlation quantity output by the prediction model, so that the confidence level decreases monotonically as the uncertainty increases; when the confidence level is less than a preset confidence level threshold, a weighted fusion is performed between a regularized linear model trained based on recent measured values ​​and previous predicted values ​​generated in the prediction sequence and the prediction mean of the prediction sequence to obtain a corrected prediction value; after the prediction model is deployed, at preset time intervals, the calibration score and empirical coverage of the prediction distribution are calculated based on recent measured values, historical predicted values ​​and variance correlation quantity; when the empirical coverage deviates from the preset ideal coverage by more than a preset allowable deviation, the confidence level threshold is adjusted within the protection interval by a preset step size.

[0149] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the industrial equipment state variable long-term domain prediction device described above can be referred to the corresponding process in the aforementioned embodiments of the industrial equipment state variable long-term domain prediction method, and will not be repeated here.

[0150] This invention also provides an electronic device for running a long-term time-domain prediction method for industrial equipment state variables; see [link to related documentation]. Figure 7 The schematic diagram of an electronic device provided by the embodiment of the present invention shown above includes a memory 400 and a processor 401. The memory 400 is used to store one or more computer instructions, which are executed by the processor 401 to implement the above-mentioned long-term domain prediction method for industrial equipment state variables.

[0151] Furthermore, Figure 7 The electronic device shown also includes a bus 402 and a communication interface 403. The processor 401, the communication interface 403 and the memory 400 are connected via the bus 402.

[0152] The memory 400 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 403 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 402 can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 7 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0153] Processor 401 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 401 or by instructions in software form. Processor 401 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a readily available storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 400, and processor 401 reads information from memory 400 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0154] This invention also provides a computer-readable storage medium storing computer-executable instructions. When these computer-executable instructions are called and executed by a processor, they cause the processor to implement the aforementioned long-term domain prediction method for industrial equipment state variables. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0155] The computer program product of the long-term domain prediction method, apparatus and electronic device for industrial equipment state variables provided in the embodiments of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods in the preceding method embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here.

[0156] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and / or device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0157] Furthermore, in the description of the embodiments of the present invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0158] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0159] 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 them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A long-time domain prediction method for state variables of industrial equipment, characterized in that, include: Historical multi-source time-series data of industrial equipment are acquired and preprocessed. Key driving variables are selected from candidate variables through dynamic response lag analysis, and a sparse graph prior matrix is ​​constructed based on the response intensity and optimal lag of the key driving variables. Based on the historical sequences of the key driving variables and the sparse graph prior matrix, the basic prediction model is trained until the preset training completion conditions are met to obtain the prediction model; wherein, the basic prediction model includes: temporal encoding, graph prior constrained attention, and three-component decoupled prediction; the loss function of the basic prediction model is based on prediction uncertainty, mean square error, three-component orthogonalization loss and weight regularization joint constraint; The real-time running data is preprocessed and then input into the trained prediction model to obtain the prediction sequence and the confidence level of each step in the prediction sequence. Online correction is performed when the confidence level is less than a preset confidence level threshold. The prediction model performs rolling calibration based on recent prediction performance after deployment.

2. The method according to claim 1, characterized in that, The sampling reference time is predetermined; the preprocessing includes: The candidate variables are interpolated and aligned at the sampling reference time. Based on the preset missing rate threshold, missing data are processed by interpolation filling, interpolation of the mean under the same working condition, or direct removal. Based on the segmented statistics of working conditions, outliers that deviate from the mean by more than a preset standard deviation are replaced by the neighborhood median. The candidate variables are normalized according to the training set statistics.

3. The method according to claim 2, characterized in that, Pre-set the lag search set; Key driver variables were selected from candidate variables through dynamic response lag analysis, including: Each lag candidate variable sequence in the lag search set is shifted and paired with the target variable sequence, and the absolute value of the Pearson correlation coefficient and the normalized mutual information based on histogram binning are calculated respectively. The absolute value and the normalized mutual information are weighted and fused by preset weights to obtain the composite response intensity. The lag that maximizes the composite response strength is taken as the optimal lag of the candidate variable, and the maximum composite response strength is taken as the response strength of the candidate variable. The mean of the response intensity of all candidate variables is used as an adaptive threshold, and variables with response intensity greater than the adaptive threshold are retained as the key driving variables.

4. The method according to claim 3, characterized in that, Based on the response strength and optimal lag of the key driving variables, a sparse graph prior matrix is ​​constructed, including: Using the target variable as the central node, the response intensity of each of the key driving variables is normalized to obtain the original weights; The original weights are corrected based on the exponential decay factor associated with the optimal lag, and the corrected weights are used as edge weights for connections between variables. Connections with edge weights less than a preset pruning threshold are removed to obtain the sparse graph prior matrix. During model training, the sparse graph prior matrix is ​​updated once every preset training period based on the statistics of the current training set.

5. The method according to claim 1, characterized in that, Timing coding, including: Extract the historical observation sequence of each of the key driving variables, input it into a temporal coding module that adopts a causal convolution, dilated causal convolution, temporal convolutional network or recurrent neural network structure, and generate a high-order representation vector; All high-order representation vectors are stacked into a multivariate feature representation matrix, and different key driving variables share the same set of encoding parameters.

6. The method according to claim 5, characterized in that, Graph prior constraints on attention include: The multivariate feature representation matrix is ​​mapped into a query matrix, a key matrix, and a value matrix through a trainable projection. The sparse graph prior matrix is ​​injected into the attention in logarithmic form, an adaptive sparsification strategy based on the number of variables is applied, and the retained attention connections are renormalized. The value matrix is ​​weighted and summed using the renormalized attention weights, and multiple layers are stacked. Residual connections and layer normalization are introduced between layers to obtain the attention output.

7. The method according to claim 6, characterized in that, Three-component decoupled prediction includes: A moving average is applied to the historical sequence of the target variable, and the resulting trend sequence is linearly extrapolated to generate trend components. The attention output of the graph prior constraint attention output is flattened and then input into a multilayer perceptron to map and generate dynamic response components. The historical residual sequence is input into a gated recurrent unit network and linearly mapped to generate residual components; wherein, the historical residual sequence is obtained by subtracting the estimated values ​​of the historical parts corresponding to the trend component and the dynamic response component from the historical sequence of the target variable; The trend component, the dynamic response component, and the residual component are summed to form a prediction sequence.

8. The method according to claim 7, characterized in that, The loss function of the basic prediction model is based on a joint constraint of prediction uncertainty, mean square error, three-component orthogonalization loss, and weight regularization, including: An independent branch is set to output the variance correlation quantity in the prediction head of the dynamic response component; The negative log-likelihood loss is calculated based on the variance correlation, and the negative log-likelihood loss, mean squared error loss, three-component orthogonalization loss, and weighted regularization term are weighted and summed to form the joint loss function of the basic prediction model.

9. The method according to claim 1, characterized in that, Online calibration, including: The confidence level of each step in the prediction sequence is calculated based on the variance correlation quantity output by the prediction model, and the confidence level decreases monotonically as the uncertainty increases. When the confidence level is less than the preset confidence level threshold, a regularized linear model trained based on recent measured values ​​and the previously generated prediction values ​​in the prediction sequence is weighted and fused with the prediction mean of the prediction sequence to obtain the corrected prediction value. Rolling calibration, including: After the prediction model is deployed, at preset time intervals, the calibration score and empirical coverage of the prediction distribution are calculated based on recent measured values, historical predicted values ​​and the variance correlation. When the empirical coverage deviates from the preset ideal coverage by more than the preset allowable deviation, the confidence base threshold is adjusted within the protection interval by a preset step size.

10. A long-time domain prediction device for the state variables of industrial equipment, characterized in that, include: The data processing module is used to acquire and preprocess historical multi-source time-series data of industrial equipment, select key driving variables from candidate variables through dynamic response lag analysis, and construct a sparse graph prior matrix based on the response strength and optimal lag of the key driving variables. The model training module is used to train the basic prediction model based on the historical sequence of the key driving variables and the sparse graph prior matrix until the preset training completion conditions are met to obtain the prediction model; wherein, the basic prediction model includes: temporal encoding, graph prior constrained attention, and three-component decoupled prediction; the loss function of the basic prediction model is based on prediction uncertainty, mean square error, three-component orthogonalization loss and weight regularization joint constraint; The data prediction module is used to input the preprocessed real-time running data into the trained prediction model to obtain the prediction sequence and the confidence level of each step in the prediction sequence, and to perform online correction when the confidence level is less than a preset confidence level threshold; wherein, the prediction model performs rolling calibration based on recent prediction performance after deployment.

Images