A boiler energy-saving performance prediction and regulation method based on industrial big data analysis

Abstract

The application discloses a kind of boiler energy-saving performance prediction and regulation method based on industrial big data analysis, comprising the following steps: S1, acquisition time series data;S2, generate standardized training dataset;S3, based on improved N-BEATS model, obtain boiler thermal efficiency prediction result;S4, identify potential anomalies in boiler thermal efficiency prediction result, automatically trigger early warning mark;S5, if thermal efficiency prediction value is lower than set threshold, then determine that current operating state has efficiency decline trend;S6, construct deep deterministic policy gradient algorithm to generate boiler operating parameters, based on thermal efficiency prediction value and boiler operating parameters Multidimensional reward function is constructed.The present application is suitable for industrial boiler energy efficiency management and intelligent optimization control scene, with the advantages of high prediction accuracy, strong periodicity identification and intelligent regulation response, can realize the advance prediction and policy energy-saving scheduling of boiler operation energy efficiency.

Description

Technical Field

[0001] This invention relates to the field of industrial process optimization control and artificial intelligence integration application technology, and in particular to a method for predicting and controlling boiler energy-saving performance based on industrial big data analysis. Background Technology

[0002] Existing boiler energy-saving optimization technologies largely rely on empirical rule-based manual adjustment methods or traditional modeling approaches. These methods primarily monitor boiler operating parameters such as fuel input, steam calorific value, and flue gas temperature, combining these with engineer experience to set thresholds and optimize adjustments. However, these methods suffer from problems such as response lag, low control precision, and inability to adapt to dynamic operating conditions, making it difficult to address the nonlinear fluctuations and periodic changes in boiler thermal efficiency under complex industrial environments. Especially in the context of multivariate coupling and the accumulation of large amounts of historical operating data, traditional rule-based or static modeling methods lack the ability to deeply mine system time-series data, making it difficult to effectively model the dynamic trend of boiler thermal efficiency evolution over time, thus limiting further exploration of energy-saving potential.

[0003] In recent years, with the integration of industrial big data and artificial intelligence technologies, deep neural networks have been gradually introduced into boiler operation modeling and performance prediction. Among them, time series prediction models represented by N-BEATS have attracted attention due to their good interpretability and residual modeling capabilities. However, the original N-BEATS model still has certain shortcomings when facing industrial boiler operation scenarios, especially in terms of insufficient ability to model periodic fluctuations, neglect of interaction characteristics between variables, and the inability of prediction results to directly guide control behavior. In addition, if the thermal efficiency prediction results cannot be fed back into the operation and control strategy in a closed loop, it can easily lead to a disconnect between prediction and control.

[0004] There is an urgent need for a method that integrates multi-dimensional time series modeling and adaptive control capabilities to accurately predict boiler thermal efficiency while constructing an intelligent feedback control mechanism based on the prediction results. This would enable proactive perception and adjustment of the boiler's operating status. Particularly in time series modeling, an improved structure with enhanced periodicity perception and cross-variable modeling capabilities is required. Furthermore, reinforcement learning algorithms should be incorporated into control decision-making to create a closed-loop linkage between prediction and control, thereby improving the overall system's energy efficiency and operational intelligence. These technical challenges are precisely the core problems that this invention aims to solve. Summary of the Invention

[0005] One objective of this invention is to propose a method for predicting and controlling boiler energy-saving performance based on industrial big data analysis. This invention integrates an improved N-BEATS time series modeling algorithm with a deep deterministic strategy gradient control method to construct a closed-loop system for predicting and controlling boiler thermal efficiency. By introducing a multi-layer feedforward neural network to extract the temporal interaction features of boiler operating parameters and extracting periodic trend features based on periodic graph convolution, accurate prediction of the evolution trend of boiler thermal efficiency is achieved. Simultaneously, the prediction results are linked with operating parameters to construct a state input vector, and a multi-dimensional reward function is introduced to drive a reinforcement learning model to continuously optimize the control strategy. This method has the advantages of high prediction accuracy, sensitivity to periodic fluctuations, and adaptive control capabilities, and can significantly improve boiler operating efficiency and energy-saving levels.

[0006] A method for predicting and controlling boiler energy-saving performance based on industrial big data analysis according to an embodiment of the present invention includes the following steps: S1. Collect time-series data during boiler operation; S2. Preprocess the time series data to generate a standardized training dataset; S3. Based on the improved N-BEATS model, the historical input window is used as training samples. Multiple blocks are stacked sequentially to model the boiler thermal efficiency time series in a residual manner. In each block, a multi-layer feedforward neural network is used to extract time interaction features, and periodic graph convolution is introduced to extract periodic trend features. The time interaction features and periodic trend features generate the prediction vector of the current block. The prediction vectors output by all blocks are summed step by step in the stacking order to obtain the boiler thermal efficiency prediction result. S4. Identify potential anomalies in boiler thermal efficiency prediction results and automatically trigger early warning indicators; S5. If the predicted thermal efficiency is lower than the set threshold, it is determined that the current operating state has a declining efficiency trend. S6. Construct a deep deterministic policy gradient algorithm to generate boiler operating parameters, and construct a multidimensional reward function based on the predicted thermal efficiency and boiler operating parameters.

[0007] Optionally, the time series data includes fuel input, air flow rate, steam production, boiler load, flue gas temperature, excess air coefficient, furnace pressure and feedwater temperature, which are aligned by time steps to construct a high-dimensional time series input sample.

[0008] Optionally, the preprocessing includes missing value imputation, outlier removal and normalization, and is divided into a historical input window and a predicted output window by a sliding window of fixed length.

[0009] Optionally, S3 specifically includes: S31. Divide the preprocessed standardized training dataset into multiple training sample pairs. Each sample pair includes a fixed-length historical input window and a corresponding prediction output window. The historical input window is used to input into the first layer Block of the improved N-BEATS model. S32. Based on the boiler operating parameters contained in the historical input window, including fuel input and steam heat generation, calculate the boiler thermal efficiency value for each time step and form a boiler thermal efficiency time series arranged in chronological order. S33. Construct an improved N-BEATS model consisting of multiple sequentially stacked Blocks. The multiple sequentially stacked Blocks use a sequential iterative residual modeling method to model the boiler thermal efficiency time series: The first Block takes a historical input window as input and outputs two sets of vectors, including a historical fitting vector and a future prediction vector. The historical fitting vector is used to reconstruct the sequence content in the historical input window, and the future prediction vector is used to predict the boiler thermal efficiency value for multiple future time steps. Starting from the second Block, the input vector of each Block is calculated by subtracting the historical fitting vector of the previous Block from the input vector of the previous Block. S34. Each block includes a time-interactive hybrid structure and a frequency domain modeling structure: The time-interactive hybrid structure performs multilayer perceptron transformation operations on the time and feature dimensions to extract time-interactive features from the boiler thermal efficiency time series. The multilayer perceptron transformation is a multilayer feedforward neural network transformation process performed on the time dimension, mainly used to extract short-term disturbance features and interactive dependencies between variables in the boiler operation time series; The periodic trend modeling structure establishes periodic connections between time steps based on the periodic graph construction method, and performs feature propagation and aggregation on the periodic graph through the graph attention mechanism to extract periodic trend features from the input vector. S35. The time interaction features and periodic trend features output by the two structures are fused to generate the historical fitting vector and future prediction vector of the current block; the future prediction vectors output by all blocks are summed step by step in stacking order to obtain the boiler thermal efficiency prediction result.

[0010] S36. Introduce a priori structural constraint of periodic trends into the Block to force the future prediction vector output by the Block to conform to a class of analyzable periodic component expression forms; the periodic component expression forms are constructed based on a combination of weighted cosine and sine functions, and by setting learnable parameters for the weight coefficients of the frequency components, the N-BEATS model is guided to explicitly model the periodic fluctuation characteristics in the boiler thermal efficiency time series.

[0011] Optionally, S4 specifically includes: S41. Based on the boiler thermal efficiency prediction results, construct a change response sequence of the boiler thermal efficiency prediction sequence. The change response sequence is constructed from the continuous time step difference of the boiler thermal efficiency prediction sequence, specifically by calculating the numerical change amplitude between adjacent boiler thermal efficiency prediction sequences in multiple future time steps. S42. Using a sliding window approach, traverse the boiler thermal efficiency prediction sequence and the change response sequence respectively. Within each sliding window, calculate the local average thermal efficiency level, local change amplitude, local standard deviation, and local change trend slope of the boiler thermal efficiency prediction sequence. At the same time, calculate the local fluctuation intensity and change rate of the change response sequence to obtain a complete set of local change characteristic indicators. S43. When the boiler thermal efficiency prediction sequence and the change response sequence change locally within a certain sliding window When the characteristic indicators exceed the set fluctuation threshold or mutation threshold, the current time window will be marked as a predicted abnormal area; S44. For time periods marked as predicted abnormal areas, generate early warning identification information and risk warning labels, and output them to the boiler operation status visualization platform.

[0012] Optionally, S5 specifically includes: S51. Perform sliding analysis on the predicted thermal efficiency values ​​within a consecutive preset length of time steps in the boiler thermal efficiency prediction sequence, and determine whether all predicted thermal efficiency values ​​within the window are lower than the set thermal efficiency threshold. S52. When the predicted thermal efficiency value within the sliding window is continuously lower than the set thermal efficiency threshold, mark the current time period as the potential efficiency decline trend range, and record its start time step and duration. S53. Analyze the boiler thermal efficiency prediction sequence marked as the efficiency decline trend interval, and combine the changes in fuel input and steam heat production within this interval to calculate the correlation index between the boiler thermal efficiency prediction sequence and fuel input and steam heat production, which is used to identify the variables that play a dominant role in the change of boiler thermal efficiency. S54. If the boiler thermal efficiency prediction sequence shows a downward trend, and the fuel input continues to rise or the steam heat output fluctuates more, it is determined that the current operating status may have problems with decreased fuel utilization or unstable steam output.

[0013] Optionally, S6 specifically includes: S61. Construct a state input vector for boiler operation control. The state input vector is composed of a combination of boiler operating parameters at the current time step. The boiler operating parameters include the predicted value of boiler thermal efficiency, fuel input, steam heat generation, flue gas temperature, excess air coefficient, and feedwater temperature. S62. Initialize the deep deterministic strategy gradient algorithm. Use a dual-network structure to construct a control action estimation network and a state value assessment network respectively, and generate boiler operating parameter adjustment actions accordingly, including the adjustment range of fuel input ratio and the adjustment range of steam heat production target. S63. Apply the boiler operating parameter adjustment action to the boiler operating parameter adjustment process and record the result after adjustment. The fuel input and steam heat production are calculated, and the difference between the actual boiler thermal efficiency feedback and the predicted boiler thermal efficiency is calculated. S64. Construct a multi-dimensional reward function to quantitatively evaluate the actual effect of boiler operating parameter adjustment; the multi-dimensional reward function is a feedback evaluation mechanism constructed by comprehensively considering the predicted value of boiler thermal efficiency, the dynamic relationship between fuel input and steam heat generation. S65. Employ an experience replay mechanism and a temporal difference update method to continuously iteratively optimize the parameters of the control action estimation network and the state value assessment network.

[0014] Optionally, the steps of the improved N-BEATS model include: In each block, the input vector first passes through a multi-layer feedforward neural network, and performs multi-layer perceptron transformation in the time dimension and feature dimension respectively to extract time interaction features; The periodic connection relationship between time steps is established based on the periodic graph construction method. The periodic trend features in the input vector are extracted by performing feature propagation and aggregation on the periodic graph through the graph attention mechanism. Starting from the second block, the input vector of each block is obtained by subtracting the historical fitting vector of the previous block from the input vector of the previous block; the future prediction vectors output by all blocks are summed step by step in stacking order to obtain the boiler thermal efficiency prediction result. Introduce prior structural constraints of periodic trends into the Block so that the prediction vector output by the Block conforms to the periodic expression form based on the weighted combination of sine and cosine.

[0015] This invention introduces an improved N-BEATS model to perform hierarchical residual modeling of boiler thermal efficiency time series. Within each block, a multi-layer feedforward neural network is used to extract time interaction features, and a periodic graph convolutional structure is combined to mine the periodic trends between time steps. This achieves high-precision, multi-scale thermal efficiency prediction of boiler operating status and effectively solves the problem that existing methods are insufficient in modeling the dynamic change trend of boiler thermal efficiency.

[0016] This invention further introduces a deep deterministic strategy gradient algorithm to construct a control strategy generation mechanism based on prediction. It uses the prediction results and current operating parameters to construct the state input, and combines a multidimensional reward function to quantitatively evaluate the energy-saving effect of the control behavior. This realizes adaptive intelligent control of boiler operating parameters and overcomes the shortcomings of traditional control methods, such as reliance on human experience, control lag, and lack of feedforward optimization capabilities.

[0017] By incorporating boiler thermal efficiency prediction error into the reward function as a feedback indicator, a linkage mechanism between prediction and control is constructed, making the reinforcement learning process sensitive to prediction bias and further improving the accuracy and stability of the control strategy. Overall, this invention achieves a closed-loop optimization process integrating prediction, identification, and control, significantly improving the energy efficiency and operational stability of boiler systems in actual industrial operation, and has promising engineering application prospects. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0019] Figure 1 This is an overall flowchart of a boiler energy-saving performance prediction and control method based on industrial big data analysis proposed in this invention. Figure 2 This is a schematic diagram of the structure of an improved N-BEATS model for predicting and controlling boiler energy-saving performance based on industrial big data analysis, as proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1-2 A method for predicting and controlling boiler energy-saving performance based on industrial big data analysis includes the following steps: S1. Collect time-series data during boiler operation; S2. Preprocess the time series data to generate a standardized training dataset; S3. Based on the improved N-BEATS model, the historical input window is used as training samples. Multiple blocks are stacked sequentially to model the boiler thermal efficiency time series in a residual manner. In each block, a multi-layer feedforward neural network is used to extract time interaction features, and periodic graph convolution is introduced to extract periodic trend features. The time interaction features and periodic trend features generate the prediction vector of the current block. The prediction vectors output by all blocks are summed step by step in the stacking order to obtain the boiler thermal efficiency prediction result. S4. Identify potential anomalies in boiler thermal efficiency prediction results and automatically trigger early warning indicators; S5. If the predicted thermal efficiency is lower than the set threshold, it is determined that the current operating state has a declining efficiency trend. S6. Construct a deep deterministic policy gradient algorithm to generate boiler operating parameters, and construct a multidimensional reward function based on the predicted thermal efficiency and boiler operating parameters.

[0022] This invention proposes a method for predicting and controlling boiler energy-saving performance based on industrial big data analysis. It comprehensively utilizes an improved N-BEATS model and a deep deterministic strategy gradient algorithm. First, it preprocesses the time-series data during boiler operation to construct standardized training samples. Then, it employs a stacked residual structure combined with time-interactive feature extraction and periodic graph convolution mechanism to achieve accurate prediction of boiler thermal efficiency. Subsequently, it identifies potential anomalies in the prediction results, triggers early warning information in a timely manner, and generates boiler operating parameter adjustment actions by constructing a multi-dimensional reward function-driven reinforcement learning mechanism when a downward trend in thermal efficiency is detected. This achieves closed-loop optimization of boiler energy-saving performance prediction, identification, and control, significantly improving operating efficiency and intelligence.

[0023] In this embodiment, the time series data includes fuel input, air flow rate, steam production, boiler load, flue gas temperature, excess air coefficient, furnace pressure and feedwater temperature, which are aligned by time steps to construct a high-dimensional time series input sample.

[0024] In this embodiment, the preprocessing includes missing value imputation, outlier removal and normalization, and is divided into a historical input window and a predicted output window by a sliding window of fixed length.

[0025] In this embodiment, S3 specifically includes: S31. Divide the preprocessed standardized training dataset into multiple training sample pairs. Each sample pair includes a fixed-length historical input window and a corresponding prediction output window. The historical input window is used to input into the first layer Block of the improved N-BEATS model. S32. Based on the boiler operating parameters contained in the historical input window, including fuel input and steam heat generation, calculate the boiler thermal efficiency value for each time step: ; in, Indicates time step Boiler thermal efficiency, This indicates the heat output of steam (in MJ). Both represent the heat energy generated by the fuel input (in the same unit), and are obtained from the corresponding time step parameters in the historical input window. The boiler thermal efficiency values ​​are arranged by time step to form a boiler thermal efficiency time series.

[0026] S33. Construct an improved N-BEATS model consisting of multiple sequentially stacked Blocks. The sequentially stacked Blocks use a sequential iterative residual modeling method to model the boiler thermal efficiency time series. The first Block uses a historical input window as input, assuming this input vector is... Output two sets of vectors: the backcast vector. and forecast vector The backcast vector is used to reconstruct the sequence content in the historical input window, and the forward vector is used to predict the boiler thermal efficiency values ​​at multiple future time steps.

[0027] Starting from the second block, the input vector of each block is calculated by subtracting the backcast vector of the previous block from the input vector of the previous block. Let the first block be... The input vector of each block is : ; in, Indicates the first The backcast vector of each block represents its residual to the input sequence. Poor reconstruction results; S34, the time-interactive hybrid structure uses the input vector of the current block. As input, a multilayer perceptron operation is first performed in the time dimension to extract the nonlinear dynamic relationships between time steps, resulting in an intermediate representation: ; Subsequently, a multilayer perceptron operation is performed along the feature dimension to extract the cross-dependencies between boiler operating variables and generate short-term perturbation feature vectors: ; in, and These represent the multilayer perceptron transformations performed in the time dimension and the feature dimension, respectively. This indicates the time step length of the historical input window. This represents the feature dimension at each time step.

[0028] Multilayer perceptron transformation in the time dimension: for the input vector In the time dimension Modeling above, taking the input vector Transpose Then, an independent multilayer perceptron transformation is performed on each row (corresponding to a boiler parameter): ; Transpose it back to get This serves as an intermediate result after time modeling.

[0029] Multilayer perceptron transformation along the feature dimension: At this time The data already contains time information, but further modeling of the cross-relationships between boiler operating parameters is needed. For each time step... ,extract That is, all boiler features at that time step are input into the multilayer perceptron transform of the feature dimension:

[0030] ; Output This represents the short-term perturbation feature representation within the Block based on time-feature dual-channel fusion.

[0031] The periodic trend modeling structure establishes periodic connections between time steps based on a periodic graph construction method. It then uses a graph attention mechanism to propagate and aggregate features on the periodic graph, extracting periodic trend features from the input vector. Step 1: Construct the adjacency matrix of the periodic graph Using the current Block's input vector as the basis for periodic trend modeling, an adjacency matrix of the periodic graph is constructed based on the boiler thermal efficiency representation at different time steps in the input vector. In this graph, if the interval between two time steps is an integer multiple of a fixed period length, it is considered that a periodic connection exists, and a corresponding connection edge is established in the adjacency matrix. The original adjacency matrix is ​​then weighted and normalized according to the node degree matrix of the graph structure to obtain a normalized periodic graph adjacency matrix.

[0032] Step 2: Periodic graph convolution Perform graph attention propagation on the normalized periodic graph, using the boiler thermal efficiency sub-vector at each time step in the input vector of the current block as the initial feature, and fuse the features of its neighboring nodes to generate a periodic trend feature vector.

[0033] The periodic trend features at each time step are represented by a weighted combination of the input vectors of its neighboring nodes. The attention weights are obtained by concatenating the features between nodes and inputting them into a ReLU activation function with leakage parameters, and then interacting with the trainable weight vectors to learn them.

[0034] Step 3: Generate a periodic trend feature matrix The periodic trend feature vectors corresponding to all time steps are concatenated to form a periodic trend feature matrix, which is used for subsequent fusion with time interaction features.

[0035] S35. The time interaction features and periodic trend features output by the two structures are fused to generate the backcast vector and forecast vector of the current block; the forecast vectors output by all blocks are summed step by step in stacking order to obtain the boiler thermal efficiency prediction result.

[0036] S36. To improve the interpretability of the Block output in the improved N-BEATS model, a priori structural constraint of periodic trends is introduced into the Block, forcing the forecast vector output by the Block to conform to an analytically reproducible periodic component expression form; specifically, the forecast vector is represented as a linear combination of a set of weighted cosine and sine functions: ; in, Indicates the first The forecast vector output by the layer block. This indicates the time step length of the prediction output window. For the current prediction time step, and The learnable frequency component weighting coefficients are used to control the contribution strength of the cosine and sine terms, respectively. The prior structural constraint refers to the periodic structural constraint imposed on a specific block in the improved N-BEATS model. This constraint restricts the output form of the forecast vector of the block, forcing it to be expressed as a linear combination of weighted sine and cosine functions. This guides the model to focus on modeling the periodic trend of boiler thermal efficiency time series in the block, thereby improving the structural transparency and interpretability of the model prediction results.

[0037] The improved N-BEATS model proposed in this invention introduces a time-feature dual-channel multilayer sensing... This method employs a knowledge-based structure and a periodic trend modeling mechanism based on periodic graph convolution to deeply mine the short-term disturbance and periodic trend features in the boiler thermal efficiency time series. Deep modeling is performed across multiple blocks using a sequential iterative residual approach, while periodic prior structural constraints are applied to specific blocks, enhancing the model's periodic modeling capability and interpretability. This approach not only improves the accuracy and stability of boiler thermal efficiency prediction but also provides reliable data support for subsequent control and optimization.

[0038] In this embodiment, S4 specifically includes: S41. Based on the boiler thermal efficiency prediction results, construct a change response sequence of the boiler thermal efficiency prediction sequence to reflect the dynamic fluctuation of boiler thermal efficiency during the prediction stage.

[0039] A change response sequence is constructed by calculating the difference between the predicted thermal efficiency values ​​between consecutive time steps. Each of them The calculation method is as follows: ; In the above formula, Indicates the first The time step and the first The absolute variation of the predicted boiler thermal efficiency between time steps reflects the intensity of the fluctuation in thermal efficiency during that period. Indicates the first Predicted boiler thermal efficiency values ​​for each prediction time step;

[0040] The changed response sequence is subjected to min-max normalization, and the normalized changed response vector is denoted as . The calculation formula is: ; in, and These represent the minimum and maximum values ​​in the change response sequence, respectively.

[0041] Normalized change response sequence It can serve as an important indicator for measuring the rate of change and the intensity of fluctuations in the boiler thermal efficiency prediction sequence, and can be used for subsequent abnormal trend identification and early warning triggering operations.

[0042] S42. Using a sliding window approach, traverse the boiler thermal efficiency prediction sequence and the change response sequence respectively. Within each sliding window, calculate the local average thermal efficiency level, local change amplitude, local standard deviation, and local change trend slope of the boiler thermal efficiency prediction sequence. At the same time, calculate the local fluctuation intensity and change rate of the change response sequence to obtain a complete set of local change characteristic indicators. The specific calculation process is as follows: Let the length of the sliding window be... The current time step index within the sliding window is

[0043] Local average thermal efficiency level for: ; Local variation amplitude for: ; Local standard deviation for: ; Local change trend slope Calculated based on the least squares method: ; Local fluctuation intensity The mean of the changing response sequence within this window: ; Local rate of change The maximum response change value for this window: ; S43. When the local change characteristic indicators corresponding to the boiler thermal efficiency prediction sequence and the change response sequence show abnormal fluctuations within a certain sliding window, the judgment operation of the prediction abnormal area is executed. Among them, the local change characteristic indicators include the local average thermal efficiency level, local change amplitude, local standard deviation and local change trend slope of the boiler thermal efficiency prediction sequence, as well as the local fluctuation intensity and change rate of the change response sequence. When any of the above indicators exceeds the preset fluctuation threshold or mutation threshold, the time period corresponding to the current sliding window is determined to be the prediction abnormal area and marked as a potential abnormal operation state area for subsequent early warning and operation intervention modules to use. S44. For time periods marked as predicted abnormal areas, generate early warning identification information and risk warning labels, and output them to the boiler operation status visualization platform.

[0044] This invention constructs a change response sequence for boiler thermal efficiency prediction and extracts multi-dimensional local change feature indicators, including local average level, change amplitude, standard deviation, trend slope, fluctuation intensity, and change rate, using a sliding window mechanism. This comprehensively reflects the dynamic change characteristics of boiler thermal efficiency during the prediction stage. When any local indicator exceeds a set fluctuation or abrupt change threshold, it is identified as an abnormal prediction area, and an early warning icon and risk alert are generated and provided to the operation status visualization platform. This enables accurate identification and intelligent early warning of abnormal trends during boiler operation. In this embodiment, S5 specifically includes: S51. Perform sliding analysis on the predicted thermal efficiency values ​​within a consecutive preset length of time steps in the boiler thermal efficiency prediction sequence, and determine whether all predicted thermal efficiency values ​​within the window are lower than the set thermal efficiency threshold. S52. When the predicted thermal efficiency value within the sliding window is continuously lower than the set thermal efficiency threshold, mark the current time period as the potential efficiency decline trend range, and record its start time step and duration. S53. Analyze the boiler thermal efficiency prediction sequence marked as the efficiency decline trend interval, and combine the changes in fuel input and steam heat production within this interval to calculate the correlation index between the boiler thermal efficiency prediction sequence and fuel input and steam heat production, which is used to identify the variables that play a dominant role in the change of boiler thermal efficiency. S54. If the boiler thermal efficiency prediction sequence shows a downward trend, and the fuel input continues to rise or the steam heat output fluctuates more, it is determined that the current operating status may have problems with decreased fuel utilization or unstable steam output.

[0045] This invention uses a sliding window analysis on the boiler thermal efficiency prediction sequence to determine whether there is a situation where the thermal efficiency is continuously lower than a set threshold within a continuous time step. This is then marked as a potential efficiency decline trend interval, and its starting position and duration are recorded. Subsequently, the correlation index between the change in fuel input and steam heat production within this interval is calculated to identify the dominant influencing factors. If the thermal efficiency shows a downward trend, accompanied by a continuous increase in fuel input or increased fluctuations in steam output, it is further determined that the boiler operation may have problems with decreased fuel utilization efficiency or unstable output, thus providing a basis for subsequent regulation.

[0046] In this embodiment, S6 specifically includes: S61. Construct a state input vector for boiler operation control. The state input vector is composed of a combination of boiler operating parameters at the current time step. The boiler operating parameters include the predicted value of boiler thermal efficiency, fuel input, steam heat generation, flue gas temperature, excess air coefficient, and feedwater temperature. S62. Initialize the deep deterministic strategy gradient algorithm. Use a dual-network structure to construct a control action estimation network and a state value assessment network respectively, and generate boiler operating parameter adjustment actions accordingly, including the adjustment range of fuel input ratio and the adjustment range of steam heat production target. S63. Apply the boiler operating parameter adjustment action to the boiler operating parameter adjustment process: Based on the boiler operating parameter adjustment actions, two components are extracted at the current time step: the fuel input ratio adjustment amplitude and the steam heat production target adjustment amplitude; the fuel input ratio adjustment amplitude and the current fuel input quantity are then compared. A weighted calculation is performed to obtain the adjusted fuel input amount, which is used to replace the original fuel input set value; the adjustment range of the steam heat production target is weighted with the current steam heat production to obtain the adjusted steam heat production target; The adjusted fuel input and steam heat production are used as new input parameters in the boiler operation state model to calculate the boiler's operating results under these conditions. After the operation ends, the actual boiler thermal efficiency feedback value generated in this operating cycle is collected, and the difference between the boiler thermal efficiency feedback value and the boiler thermal efficiency prediction value generated by the improved N-BEATS model at the same time step is calculated. This difference is used for the subsequent construction of the reinforcement learning reward function and the parameter update operation of the control network.

[0047] S64. Construct a multidimensional reward function to quantitatively evaluate the actual effect of boiler operating parameter adjustment; the multidimensional reward function is a feedback evaluation mechanism constructed by comprehensively considering the predicted value of boiler thermal efficiency, the dynamic relationship between fuel input and steam heat generation.

[0048] S65. Employing an experience replay mechanism and a temporal difference update method, the parameters of the control action estimation network and the state value assessment network are continuously and iteratively optimized; the details are as follows: After each boiler operating parameter adjustment is completed, the current time step's state input vector, the generated boiler operating parameter adjustment action, the adjusted actual boiler thermal efficiency feedback value, the next time step's state input vector, and the reward value quintuple output by the multidimensional reward function are packaged and stored in the experience playback cache pool to form an interaction record sample between the boiler operating state and the adjustment behavior.

[0049] During the training phase of the control network, multiple interaction samples are randomly sampled in mini-batch from the experience replay buffer. For each sampled sample, the state value estimate corresponding to the current state input vector is calculated. Based on the state input vector of the next time step, the control action estimation network predicts the next adjustment action. The action and the next state input vector are then input into the state value evaluation network to generate the target state value.

[0050] Based on the temporal difference method, the difference between the current state value estimation result and the target state value is taken as the temporal difference error. Minimizing this error is taken as the optimization objective, and the parameters of the state value assessment network are updated. The state value assessment network is frozen, and its output value is fixed as the objective. It is backpropagated to the control action estimation network to update its parameters, so that the boiler operating parameter adjustment action generated in the current state can obtain better state feedback and reward value in the next time step, thereby achieving continuous convergence and performance improvement of the overall control strategy.

[0051] This invention constructs a state input vector containing predicted boiler thermal efficiency and various operating parameters, and combines it with a deep deterministic strategy gradient algorithm to establish a boiler operation and control model composed of a control action estimation network and a state value assessment network. This model generates adjustment actions based on the fuel input ratio and steam heat production target, and applies them to the boiler operating state model, obtaining real-time feedback on actual thermal efficiency. By constructing a multidimensional reward function that combines the relationship between predicted values ​​and changes in operating parameters, the adjustment effect is quantitatively evaluated. Furthermore, the control network parameters are continuously optimized using an experience playback mechanism and a time-series difference method, thereby achieving intelligent adjustment of boiler operating parameters and dynamic adaptive updating of the control strategy, significantly improving operating efficiency and response flexibility.

[0052] Example 1: To verify the application effect of this invention in a real engineering environment, this paper takes the operation scenario of a chain grate boiler in an industrial heating station as an example and conducts an experiment on energy-saving performance prediction and control based on industrial big data analysis. This heating station is equipped with multiple chain grate boilers, possessing continuous and stable coal-fired heating capacity. Parameters collected during operation include fuel input, steam heat production, flue gas temperature, excess air coefficient, feedwater temperature, and boiler thermal efficiency. The data sampling period is 10 minutes, with a continuous collection period covering 30 days, and the total data volume exceeds 40,000 records. After preprocessing and standardization, the data is input into the improved N-BEATS model proposed in this invention for thermal efficiency prediction analysis and efficiency decline trend identification. The table below shows the key performance data indicators of this invention in practical applications.

[0053] Table 1 Comparison of key boiler operation indicators after the application of this invention

[0054] In the experiment, a fixed-length historical input window (set to 24 time steps, corresponding to 4 hours of operating data) was first constructed to extract the dynamic evolution characteristics of boiler operating variables. The improved N-BEATS model introduced a time-interactive hybrid structure in each block to identify the cross-dependencies between boiler operating variables. Simultaneously, a periodic graph convolution structure was fused to extract the implicit periodic trends in boiler operation, significantly improving the modeling ability for energy-saving characteristics. Prediction results show that in the 24-step rolling prediction task of boiler thermal efficiency, the model has a mean absolute error (MAE) of 1.83% and a relative error (MAPE) of no more than 3.25%, representing reductions of 38.2% and 42.5% respectively compared to the traditional LSTM model.

[0055] Furthermore, by using a sliding window to analyze local changes and detect anomalies in the predicted thermal efficiency sequence, several potential efficiency decline trend segments were successfully identified. For example, from 14:00 to 18:00 on day 12 of the experiment, the predicted sequence was continuously below the thermal efficiency benchmark threshold (78%). Sliding window analysis showed that the steam heat production fluctuations intensified and the flue gas temperature continued to rise in this segment, initially indicating a possible decline in fuel utilization. Based on this identification result, the system automatically activated the reinforcement learning control module, using the Deep Deterministic Policy Gradient Algorithm (DDPG) to generate operating parameter adjustment actions, mainly focusing on the joint optimization of fuel input ratio and steam heat production target. Within two hours after the control action was implemented, the predicted thermal efficiency rebounded from 76.1% to 80.4%, and the actual feedback thermal efficiency reached 81.2%, demonstrating significant effectiveness in the control response.

[0056] During the reinforcement learning-based control training process, the model employs a five-tuple experience replay mechanism and a temporal differential update strategy to achieve stable optimization of the control strategy. The training process lasted approximately 48 hours, accumulating over 200,000 state-action samples. After strategy convergence, it can stably output operating parameters with energy-saving optimization effects. Within a 30-day operating cycle at this thermal power station, after implementing the predictive-control closed-loop system of this invention, the average thermal efficiency increased by 2.6% compared to the baseline operating mode, and the unit fuel energy consumption decreased by approximately 1.8%, with an estimated annual fuel cost saving of approximately 270,000 yuan (based on current fuel prices).

[0057] As can be seen from the above experimental results, this invention not only significantly improves the accuracy of boiler thermal efficiency prediction, but also achieves precise adjustment of operating parameters by integrating reinforcement learning algorithms, effectively suppressing efficiency fluctuations and abnormal trends, and possesses good engineering practicality and energy-saving economic value.

[0058] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting and controlling boiler energy-saving performance based on industrial big data analysis, characterized in that, Includes the following steps: S1. Collect time-series data during boiler operation; S2. Preprocess the time series data to generate a standardized training dataset; S3. Based on the improved N-BEATS model, the historical input window is constructed as training samples, and multiple blocks are stacked in the residual manner to obtain the boiler thermal efficiency time series. In each block, a multi-layer feedforward neural network is used to extract time interaction features, and periodic graph convolution is introduced to extract periodic trend features. The time interaction features and periodic trend features generate the historical fitting vector and future prediction vector of the current block to obtain the boiler thermal efficiency prediction results. S3 specifically includes: S31. Divide the preprocessed standardized training dataset into multiple training sample pairs. Each sample pair includes a fixed-length historical input window and a corresponding prediction output window. The historical input window is used to input into the first layer Block of the improved N-BEATS model. S32. Based on the boiler operating parameters contained in the historical input window, including fuel input and steam heat generation, calculate the boiler thermal efficiency value for each time step and form a boiler thermal efficiency time series arranged in chronological order. S33. Construct an improved N-BEATS model consisting of multiple sequentially stacked Blocks. The multiple sequentially stacked Blocks use a sequential iterative residual modeling method to model the boiler thermal efficiency time series: The first Block takes a historical input window as input and outputs two sets of vectors, including a historical fitting vector and a future prediction vector. The historical fitting vector is used to reconstruct the sequence content in the historical input window, and the future prediction vector is used to predict the boiler thermal efficiency value for multiple future time steps. Starting from the second Block, the input vector of each Block is calculated by subtracting the historical fitting vector of the previous Block from the input vector of the previous Block. S34. Each block includes a time-interactive hybrid structure and a frequency domain modeling structure: The time-interactive hybrid structure performs multilayer perceptron transformation operations on the time and feature dimensions to extract time-interactive features from the boiler thermal efficiency time series. The multilayer perceptron transformation is a multilayer feedforward neural network transformation process performed on the time dimension, mainly used to extract short-term disturbance features and interactive dependencies between variables in the boiler operation time series; The periodic trend modeling structure establishes periodic connections between time steps based on the periodic graph convolution construction method, and performs feature propagation and aggregation on the periodic graph through the graph attention mechanism to extract periodic trend features from the input vector. S35. The time interaction features and periodic trend features of the two structural outputs are fused to generate the historical fitting vector and future prediction vector of the current block; the future prediction vectors of all blocks are summed step by step in stacking order to obtain the boiler thermal efficiency prediction result. S36. Introduce a priori structural constraint of periodic trends into the Block to force the future prediction vector output by the Block to conform to a class of analyzable periodic component expression forms; the periodic component expression forms are constructed based on a combination of weighted cosine and sine functions, and by setting learnable parameters for the weight coefficients of the frequency components, the N-BEATS model is guided to explicitly model the periodic fluctuation characteristics in the boiler thermal efficiency time series. S4. Identify potential anomalies in boiler thermal efficiency prediction results and automatically trigger early warning indicators; S5. If the predicted thermal efficiency is lower than the set threshold, it is determined that the current operating state has a declining efficiency trend. S6. Construct a deep deterministic policy gradient algorithm to generate boiler operating parameters, and construct a multidimensional reward function based on the predicted thermal efficiency and boiler operating parameters.

2. The method for predicting and controlling boiler energy-saving performance based on industrial big data analysis according to claim 1, characterized in that, The time series data includes fuel input, air flow rate, steam production, boiler load, flue gas temperature, excess air coefficient, furnace pressure and feedwater temperature, which are aligned by time steps to construct a high-dimensional time series input sample.

3. The method for predicting and controlling boiler energy-saving performance based on industrial big data analysis according to claim 1, characterized in that, The preprocessing includes missing value imputation, outlier removal and normalization, and is divided into a historical input window and a predicted output window by a sliding window of fixed length.

4. The method for predicting and controlling boiler energy-saving performance based on industrial big data analysis according to claim 1, characterized in that, S4 specifically includes: S41. Based on the boiler thermal efficiency prediction results, construct a change response sequence of the boiler thermal efficiency prediction sequence. The change response sequence is constructed from the continuous time step difference of the boiler thermal efficiency prediction sequence, specifically by calculating the numerical change amplitude between adjacent boiler thermal efficiency prediction sequences in multiple future time steps. S42. Using a sliding window approach, traverse the boiler thermal efficiency prediction sequence and the change response sequence respectively. Within each sliding window, calculate the local average thermal efficiency level, local change amplitude, local standard deviation, and local change trend slope of the boiler thermal efficiency prediction sequence. At the same time, calculate the local fluctuation intensity and change rate of the change response sequence to obtain a complete set of local change characteristic indicators. S43. When the boiler thermal efficiency prediction sequence and the change response sequence change locally within a certain sliding window When the characteristic indicators exceed the set fluctuation threshold or mutation threshold, the current time window will be marked as a predicted abnormal area; S44. For time periods marked as predicted abnormal areas, generate early warning identification information and risk warning labels, and output them to the boiler operation status visualization platform.

5. The method for predicting and controlling boiler energy-saving performance based on industrial big data analysis according to claim 1, characterized in that, S5 specifically includes: S51. Perform sliding analysis on the predicted thermal efficiency values ​​within a consecutive preset length of time steps in the boiler thermal efficiency prediction sequence, and determine whether all predicted thermal efficiency values ​​within the window are lower than the set thermal efficiency threshold. S52. When the predicted thermal efficiency value within the sliding window is continuously lower than the set thermal efficiency threshold, mark the current time period as the potential efficiency decline trend range, and record the start time step and duration. S53. Analyze the boiler thermal efficiency prediction sequence marked as the efficiency decline trend interval, and calculate the correlation index between the boiler thermal efficiency prediction sequence and the fuel input and steam heat production in combination with the change direction of fuel input and steam heat production in the interval, so as to identify the variables that play a dominant role in the change of boiler thermal efficiency. S54. If the boiler thermal efficiency prediction sequence shows a downward trend, and the fuel input continues to rise or the steam heat output fluctuates more, it is determined that the current operating status may have problems with decreased fuel utilization or unstable steam output.

6. The method for predicting and controlling boiler energy-saving performance based on industrial big data analysis according to claim 1, characterized in that, S6 specifically includes: S61. Construct a state input vector for boiler operation control. The state input vector is composed of a combination of boiler operating parameters at the current time step. The boiler operating parameters include the predicted value of boiler thermal efficiency, fuel input, steam heat generation, flue gas temperature, excess air coefficient, and feedwater temperature. S62. Initialize the deep deterministic strategy gradient algorithm. Use a dual-network structure to construct a control action estimation network and a state value assessment network respectively, and generate boiler operating parameter adjustment actions accordingly, including the adjustment range of fuel input ratio and the adjustment range of steam heat production target. S63. Apply the boiler operating parameter adjustment action to the boiler operating parameter adjustment process, record the fuel input and steam heat generation after adjustment, and calculate the difference between the actual boiler thermal efficiency feedback and the predicted boiler thermal efficiency. S64. Construct a multi-dimensional reward function to quantitatively evaluate the actual effect of boiler operating parameter adjustment; the multi-dimensional reward function is a feedback evaluation mechanism constructed by comprehensively considering the predicted value of boiler thermal efficiency, the dynamic relationship between fuel input and steam heat generation. S65. Employ an experience replay mechanism and a temporal difference update method to continuously iteratively optimize the parameters of the control action estimation network and the state value assessment network.

7. The method for predicting and controlling boiler energy-saving performance based on industrial big data analysis according to claim 1, characterized in that, The steps of the improved N-BEATS model include: In each block, the input vector first passes through a multi-layer feedforward neural network, and performs multi-layer perceptron transformation in the time dimension and feature dimension respectively to extract time interaction features; The input vector establishes periodic connections between time steps based on a periodic graph construction method. Feature propagation and aggregation are performed on the periodic graph through a graph attention mechanism to extract periodic trend features from the input vector. Starting from the second block, the input vector of each block is obtained by subtracting the historical fitted vector from the input vector of the previous block; the future prediction vectors output by all blocks are summed step by step in stacking order to obtain the boiler thermal efficiency prediction result. A priori structural constraint of periodic trends is introduced into the Block, so that the prediction vector output by the Block conforms to the periodic expression form based on the weighted combination of sine and cosine.

Images