A prediction method based on adaptive boiler thermal energy

Abstract

The present invention relates to the technical field of industrial boilers, and provides a prediction method based on adaptive boiler thermal energy. It includes data collection and preprocessing, model selection and training, adaptive adjustment, prediction and evaluation stages. Data collection covers multiple types of parameters, environmental data and historical data, and corrects abnormal values; model selection and training adopts a multi-model fusion strategy to build a hybrid model; adaptive adjustment uses a variety of algorithms to dynamically adjust model parameters and structures through real-time data monitoring and working condition identification; the prediction and evaluation stage combines the working condition optimization model fusion prediction, calculates the comprehensive performance evaluation index, and visualizes and analyzes the results. The present invention improves the prediction accuracy and stability, helps to optimize energy management, reduce costs, and ensure production safety.

Description

Technical Field

[0001] The invention relates to the technical field of industrial boilers, and in particular to a prediction method based on adaptive boiler heat energy. Background Art

[0002] With the rapid development of industry, boilers, as important thermal energy supply equipment, play a key role in many industries such as electricity, chemical industry, and papermaking. However, the boiler system is highly complex and nonlinear, and its thermal energy output is affected by the interaction of multiple factors, including the quality and flow of fuel, the flow and temperature of water, the amount of air supplied during the combustion process, the load change of the boiler, and the operating status of the equipment itself. Traditional thermal energy prediction methods based on physical models or single empirical models are difficult to accurately capture these complex relationships, resulting in large deviations in the prediction results.

[0003] In actual operation, boilers often face frequent changes in operating conditions, such as starting, stopping, load adjustment, and fuel switching. These changes in operating conditions will cause dramatic fluctuations in the system's internal parameters, further increasing the difficulty of thermal energy prediction. For example, during the boiler startup phase, the temperature and pressure of each component gradually rise, and the system is in an unstable state. Traditional models often cannot accurately predict the thermal energy output during this period; and when the load changes, the fuel-air ratio needs to be adjusted accordingly. If the thermal energy changes cannot be accurately predicted in a timely manner, it may lead to energy waste or insufficient heating.

[0004] In addition, with the continuous rise in energy costs and increasingly stringent environmental protection requirements, industrial enterprises have put forward higher requirements for the efficient operation and precise control of boilers. Accurate thermal energy prediction not only helps to optimize the operating parameters of boilers, improve energy efficiency, and reduce operating costs, but also provides an important basis for preventing equipment failures in advance and ensuring production safety. Therefore, it is of great practical significance to develop a boiler thermal energy prediction technology that can adapt to complex operating conditions and has high accuracy and stability. Summary of the invention

[0005] The purpose of this section is to summarize some aspects of embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions will be made in this section and the specification abstract and the invention title of this application to avoid blurring the purpose of this section, the specification abstract and the invention title, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0006] In view of the above problems and / or the problems existing in the prior art, the present invention is proposed.

[0007] Therefore, the purpose of the present invention is to overcome the deficiencies in the prior art and provide a prediction method based on adaptive boiler thermal energy, which comprises the following steps:

[0008] S100, data collection and preprocessing stage;

[0009] S200, model selection and training phase;

[0010] S300, adaptive adjustment stage;

[0011] S400, prediction and evaluation stage;

[0012] As a preferred solution of the present invention, S100, the data collection stage includes the following steps:

[0013] S110, parameter collection;

[0014] S120, environmental data collection;

[0015] S130, historical data collection;

[0016] S140, outlier detection and correction;

[0017] The basic parameters of the boiler in the S100, S110, parameter collection in the data collection phase include boiler model, capacity, design thermal efficiency, heating area, etc. These parameters determine the basic performance and potential heat energy generation capacity of the boiler.

[0018] Fuel-related data: including fuel type (such as coal, natural gas, biomass, etc.), fuel calorific value (the amount of heat released by complete combustion of fuel per unit mass or volume), and fuel supply rate. Fuel calorific value is the key to calculating heat input. Assume that the fuel calorific value is q (k / kg or k / m 3 ), the fuel supply rate is m (kg / s or m 3 / s), then the heat provided by the fuel is Q fuel =q×m.

[0019] Operation parameters: record the boiler inlet water temperature T in , water outlet temperature T out . Water flow rate V, pressure (steam pressure or hot water pressure) P, burner power (gas or oil boiler) W, etc. The inlet water temperature and water flow rate determine the amount of water to be heated and the initial state. The outlet water temperature reflects the heat transfer effect. According to the heat calculation formula Q = c × m × ΔT (where c is the specific heat capacity of water, m is the mass of water, and ΔT is the temperature change), the heat absorbed or released by water can be calculated, which is then associated with the boiler heat energy.

[0020] S100, S120 of the data collection phase, and the ambient temperature T during the environmental data collection env , and humidity H env Data: Ambient temperature affects boiler heat dissipation loss. According to the heat conduction formula Q loss= k × A × (T bolier -T env )(where k is the heat transfer coefficient, A is the heat dissipation area, T bolier = is the boiler surface temperature). It can be seen that the higher the ambient temperature, the smaller the heat loss (assuming that the boiler surface temperature is higher than the ambient temperature). Humidity has a certain impact on the combustion process in some biomass combustion situations. High humidity environment reduces the combustion efficiency of biomass fuel. These data are obtained through meteorological sensors installed near the boiler room so that the impact of environmental factors on boiler heat energy can be considered in subsequent analysis.

[0021] In the above S100, S130 of the data collection phase, the historical data collection obtains the operating data of the past period from the boiler control system or data recording system, including the historical records of the above parameters under different working conditions, to provide a data basis for subsequent model training and adaptive adjustment. Assume that the inlet water temperature at the i-th moment in the historical data is T in (i), the outlet water temperature is T out (i), water flow rate is V(i), etc. These data are organized into time series data for analyzing the thermal energy variation law of the boiler under different operating conditions. By analyzing the parameter changes under different loads in the historical data, the characteristic patterns related to thermal energy output are found.

[0022] S100, S140 of the data collection phase, outlier detection and correction: use the density-based local outlier factor (LOF) algorithm to detect outliers. Let the adjacent normal data points be T in (i-1) and T in (i+1), the time weights are w1 and w2 (satisfying w1+w2=1 and w1>w2), then the corrected inlet water temperature

[0023] As a preferred solution of the present invention, S200, the model selection and training stage includes the following steps:

[0024] S210, hybrid model construction;

[0025] S220, multi-model fusion strategy;

[0026] S230, a method for determining a model fusion weight;

[0027] In the S200, model selection and training phase, S210, hybrid model construction: In the boiler thermal energy system, some factors have a linear relationship with the thermal energy output, while some have complex nonlinear characteristics. Under stable operation, there is a relatively direct linear correlation between water flow and thermal energy output, and the relationship is expressed as: linear =β0+β1x1+…+β n x n

[0028] Among them, y linear represents the thermal energy predicted by the linear part, x i Represents linearly related input parameters (such as water flow during stable operation, etc.) β i is the corresponding coefficient. Coefficient β i It can be estimated by the least squares method, and its calculation formula is:

[0029] β=(X T X) -1 X T y

[0030] Among them, X is the input parameter matrix (each row represents a set of input parameter samples, and each column corresponds to an input parameter), and y is the actual thermal energy output vector.

[0031] Assume that the input layer has n neurons (corresponding to n input parameters) and the jth hidden layer has m j neurons, and the output layer has 1 neuron (for predicting thermal energy). For the i-th neuron in the l-th layer (l=2,…,L, L is the total number of layers), its input With output The relationship is as follows:

[0032]

[0033]

[0034] in, is the connection weight from the jth neuron in the (l-1)th layer to the ith neuron in the lth layer, is the bias of the ith neuron in the lth layer, and f(·) is the activation function (such as the ReLU function: f(x) = max(0,x)).

[0035] The final hybrid model prediction result is y = y linear +y neural In order to optimize the fusion effect, the weights of the two models can be dynamically adjusted according to their performance under different working conditions. Suppose the performance evaluation value of the linear model under the current working condition is p linear (It can be calculated based on mean square error, accuracy and other indicators). The performance evaluation value of the neural network model under the current working conditions is p neural , then the weight w of the linear model linear , and the weights w of the neural network model neural It can be calculated by the following formula:

[0036]

[0037] Under different operating conditions, the model weights can be automatically adjusted according to the performance, making the hybrid model better suited to different stages of boiler operation.

[0038] The linear regression model selected in the S200, model selection and training phase, and the multi-model fusion strategy is more accurate in describing the linear relationship and can provide a relatively stable prediction benchmark. The model formula is:

[0039] y LR =β0+β1x1+…+β n x n , the support vector machine model has good generalization ability when dealing with small samples and nonlinear problems. Its prediction formula is Where N is the number of support vectors, α i is the Lagrange multiplier, yi is the sample label, K(xi,x) is the kernel function (such as the radial basis kernel function K(xi,x) = exp(-γxi-x 2 ), γ is the kernel function parameter). In boiler thermal energy prediction, when the data samples are limited and there are complex nonlinear relationships, support vector machines can classify or regress data by finding the optimal hyperplane, and can achieve better results for data sets with complex boundaries.

[0040] S200, S230 in the model selection and training phase, and the method for determining the weight of the model fusion weight based on the comprehensive evaluation index Assume w i (t) is the model M i The weight at time t, e i (t) is the model M i The performance indicator (such as mean square error) calculated based on the latest data at time t, the weight update formula is

[0041] As a preferred solution of the present invention, the adaptive adjustment stage S300 includes the following steps:

[0042] S310, real-time data monitoring and working condition identification;

[0043] S320, dynamic adjustment of model parameters;

[0044] S330, adaptive optimization of model structure;

[0045] In the S300, S310 of the adaptive adjustment stage, the real-time data monitoring and working condition identification adopts the working condition identification based on wavelet transform and hidden Markov model (HMM): the real-time data of boiler operation (such as temperature, pressure, flow, etc.) is subjected to wavelet transform. Assume that the real-time data sequence is x(n), n=1,2,…,N, and the wavelet transform formula is Where a is the scale factor, b is the translation factor, and ψ(·) is the wavelet basis function. The data is decomposed into sub-bands of different frequencies through wavelet transform, and the energy characteristics E of the sub-bands are extracted. i ,i=1,2,…,k (k is the number of subbands). These energy features constitute the feature vector E=[E1,E2,…,E k ] T .

[0046] These feature vectors are used as the observation sequence of the Hidden Markov Model (HMM). The HMM consists of hidden states S = {s1, s2, ..., s M} (M is the number of states, such as startup, stable operation, variable load and other working conditions), observation symbol O = {o1, o2, ..., o L} (here the observation symbol can correspond to different energy characteristic value ranges), initial state probability distribution π = (π1, π2, ... π n )(π i =P(q1=s i ), q1 is the initial state), state transition probability matrix

[0047] A=(a ij )(a ij =P(q t+1 =s j q t =s j ),t=1,2,…) and the observation probability matrix B=(b ij )(b ij =P(o t =o j q t =s j ))composition.

[0048] The Adadelta optimization algorithm is used in the dynamic adjustment of model parameters in the S300 and S320 of the adaptive adjustment stage: During the model training process, the Adadelta optimization algorithm dynamically adjusts the learning rate by accumulating the square of the past gradient. Assume that the gradient of parameter θ at time t is g t , first calculate the exponential moving average of the square of the gradient (ρ is the reduction rate, usually between 0 and 1, such as 0.95).

[0049] Then calculate the parameter update amount Where η is the initial learning rate and ε is a minimum value (such as 10 -6 ) to prevent the denominator from being zero, E[θ 2 ] t-1 It is also an exponential moving average, and its update formula is The Adadelta optimization algorithm can adaptively adjust the learning rate based on the historical information of the gradient. When processing noisy and dynamically changing data such as boiler operation data, it can make the model parameter update more stable and effective, and improve the model's adaptability to new data.

[0050] In the S300 and S330 stages of the adaptive adjustment, the model structure adaptive optimization adopts a neural network structure adjustment based on particle swarm optimization (PSO): the neural network structure parameter vector is X=[x1, x2, ...x D ] T (D is the structural parameter dimension, such as: x1 is the number of hidden layers, x2 is the number of neurons in the first layer, etc.).

[0051] Define the particle's velocity vector V = [v1, v2, ... v D ] T Each particle in the particle swarm updates its speed and position according to its own historical optimal position pbest and the global optimal position gbest of the swarm. The speed update formula is v i,d (t+1)=wv i,d (t)+c1r1(pbest i,d -x i,d (t))+c2r2(gbest d -x i,d (t)), where w is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers between [0,1], and t is the number of iterations. The position update formula is x i,d (t+1)=v i,d (t)+v i,d (t+1).

[0052] Substitute the neural network structure parameters into the model, calculate the performance index of the model on the validation set, and use the performance index as the fitness function of the particle. Through multiple generations, the particle swarm continuously searches for the optimal neural network structure, so that the model can automatically adjust the structure according to the changes in the boiler operating conditions, thereby improving the prediction performance and adaptability of the model.

[0053] As a preferred solution of the present invention, in the prediction and evaluation stage S400, after the adaptive adjustment stage (S300), the model has optimized its own structure and parameters according to the real-time monitoring data and the working condition identification results, laying a solid foundation for accurate thermal energy output prediction. It includes the following steps:

[0054] S410, heat output prediction;

[0055] S420, performance evaluation index calculation;

[0056] S430, Result visualization and analysis;

[0057] In the S400, prediction and evaluation phase, S410, heat output prediction, a prediction model fusion combining working condition adaptation is performed, and w i,k is the initial weight of model i under working condition k, which is determined by normalizing the inverse of the model performance evaluation index (such as mean square error) under the same historical working condition. As new data continues to flow in, the prediction performance of each model under the current working condition is updated in real time. The sliding window technology is used to calculate the performance index of the most recent p time steps (p can be set according to the dynamic characteristics of the data) to dynamically adjust the model weight. Assume that the predicted value of model i at time t under working condition k is The fusion prediction value (m is the total number of models).

[0058] Assume that the source condition model parameter is θ s , the target operating condition model parameter is θ t , by calculating the similarity sim between the source working condition and the target working condition data (which can be calculated based on data distribution characteristics or feature correlation), the target working condition model parameters are updated by weighted average: (α is a weight parameter, which is adjusted according to the actual situation). On this basis, the model is fine-tuned and trained using the enhanced target working condition data to obtain a prediction model suitable for the current working condition, thereby improving the prediction ability under different working conditions, especially new working conditions or data-scarce working conditions.

[0059] The S400, S420 of the prediction and evaluation phase, the dynamic evaluation index considering the change of working conditions in the performance evaluation index calculation, the comprehensive performance evaluation index (β, γ, δ are weight coefficients, which are determined according to the importance of the impact of each part under different working conditions).

[0060] The degree of probability distribution matching between the two is calculated as the performance evaluation index. Assume that the probability distribution of the predicted value is P y , the probability distribution of the actual value is Q y , taking KL divergence as an example, the evaluation index The smaller the index is, the closer the probability distribution of the predicted value is to the actual value, and the better the model performance is.

[0061] S430 Result Visualization and Analysis in the S400 Prediction and Evaluation Phase: Create a visualization interface to associate prediction results with operating condition changes, and display the heat output prediction, actual measurement value curve, and operating condition switching mark points on the time axis simultaneously. Use color or line type to distinguish different operating condition data segments, and display the dynamic changes of operating condition-related parameters. Interactive operations can pop up detailed information. Decompose the prediction error components (using the time series decomposition method), and display the error components over time in the interface with stacked bar charts or line charts. Analyze the relationship between trend errors and long-term boiler operation trends, and the correlation between period errors and operating cycles, to help users identify error sources and optimize model or boiler operation.

[0062] The present invention provides a prediction method based on adaptive boiler heat energy, which has the following beneficial effects:

[0063] 1. Improve prediction accuracy: The present invention adopts a multi-model fusion strategy, which combines the advantages of various types of models such as linear regression models, support vector machine models and decision tree models. By dynamically determining the fusion weight according to the performance of each model under different working conditions, it is possible to more accurately capture the complex relationship between the boiler's thermal output and various input parameters, effectively reducing the prediction error caused by the limitations of a single model, thereby significantly improving the accuracy of the overall prediction. For example, when dealing with the nonlinear relationship between fuel composition and thermal output, the support vector machine model can play its advantages; and for the linear part of water flow and thermal output, the linear regression model can provide stable and reliable predictions. The combination of the two and the decision tree model's judgment and adjustment of the working conditions make the prediction results closer to the actual value.

[0064] 2. Enhanced adaptability and stability: In the adaptive adjustment stage, through real-time data monitoring and working condition identification, multi-scale sliding window data processing and working condition identification model based on deep learning can timely and accurately judge the operating condition of the boiler. Based on this, the model parameters are dynamically adjusted, such as adaptive learning rate adjustment based on gradient descent and parameter regularization and sparsification adjustment, and the model structure is adaptively optimized, such as model structure search based on genetic algorithm and dynamic neuron increase and decrease strategy and other innovative methods, so that the model can automatically adjust its own parameters and structure according to the changes in working conditions, so as to better adapt to various complex working condition changes during the operation of the boiler, effectively avoid the decline of model performance due to working condition switching or data fluctuations, and greatly enhance the adaptability and stability of the prediction model.

[0065] 3. Optimize energy management and reduce costs: Accurate thermal energy prediction results provide strong support for boiler energy management. Enterprises can adjust operating parameters such as fuel supply and water flow in advance based on the forecast information to achieve reasonable allocation and efficient use of energy and avoid excessive consumption or waste of energy. For example, when the load demand is low, timely reduce fuel input to reduce operating costs. At the same time, through accurate prediction of the operating status of the boiler, equipment maintenance and overhaul plans can be arranged in advance to reduce sudden downtime caused by equipment failure, further improve production efficiency, and reduce maintenance costs and economic losses caused by downtime.

[0066] 4. Ensure production safety and reliability: Since the present invention can monitor the boiler operating conditions in real time and accurately predict the heat output, it helps to timely discover potential abnormal conditions, such as abnormal changes in heat energy that may be caused by insufficient combustion and pipeline blockage. Through the early warning mechanism, operators can take measures to intervene in advance, effectively prevent accidents, ensure the safety and reliability of boiler operation, and provide solid guarantee for the stable progress of industrial production. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0068] Figure 1 A schematic diagram of a system framework in one embodiment of the present invention;

[0069] Figure 2 A schematic diagram of a method flow in one embodiment of the present invention;

[0070] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION

[0071] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the present invention. It should be understood that the specific implementation methods described here are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0072] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the embodiments of the specification.

[0073] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0074] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0075] As a preferred embodiment of the present invention, the present invention provides a prediction method based on adaptive boiler thermal energy, which comprises the following steps:

[0076] S100, data collection and preprocessing stage;

[0077] S200, model selection and training phase;

[0078] S300, adaptive adjustment stage;

[0079] S400, prediction and evaluation stage;

[0080] As a preferred solution of the present invention, S100, the data collection stage includes the following steps:

[0081] S110, parameter collection;

[0082] S120, environmental data collection;

[0083] S130, historical data collection;

[0084] S140, outlier detection and correction;

[0085] The basic parameters of the boiler in the S100, S110, parameter collection in the data collection phase: covering boiler model, capacity, design thermal efficiency, heating area, etc. These parameters determine the basic performance and potential heat energy generation capacity of the boiler. The boiler model is HG-2980 / 26.15-YM2 single reheat ultra-supercritical variable pressure operation direct current boiler. Its design parameters reflect the specific heat conversion capacity and working range, expressed as: design (design pressure), T design (Design temperature), Q design (design thermal power), etc. These parameters act as basic constants in subsequent calculations and model building to affect the prediction results.

[0086] Fuel-related data: including fuel type (such as coal, natural gas, biomass, etc.), fuel calorific value (the amount of heat released by complete combustion of fuel per unit mass or volume), and fuel supply rate. Fuel calorific value is the key to calculating heat input. Assume that the fuel calorific value is q (k / kg or k / m 3 ), the fuel supply rate is m (kg / s or m 3 / s), then the heat provided by the fuel is Q fuel =q×m. The calorific value of different fuels varies significantly. For example, the calorific value of coal is generally lower than that of natural gas. Changes in the fuel supply rate directly affect the amount of heat generated by combustion, which in turn affects the heat output of the boiler.

[0087] Operation parameters: record the boiler inlet water temperature T in , water outlet temperature T out . Water flow rate V, pressure (steam pressure or hot water pressure) P, burner power (gas or oil boiler) W, etc. The inlet water temperature and water flow rate determine the amount of water to be heated and the initial state. The outlet water temperature reflects the heat transfer effect. According to the heat calculation formula Q = c × m × ΔT (where c is the specific heat capacity of water, m is the mass of water, and ΔT is the temperature change), the heat absorbed or released by water can be calculated, and then related to the boiler heat energy. Pressure changes are related to steam generation and transportation. The burner power is directly related to the fuel combustion intensity and the heat energy release rate. For example, the greater the burner power, the more energy released by fuel combustion per unit time, which has an important impact on the boiler heat energy.

[0088] S100, S120 of the data collection phase, and the ambient temperature T during the environmental data collection env , and humidity H env Data: Ambient temperature affects boiler heat dissipation loss. According to the heat conduction formula Q loss = k × A × (T bolier -T env )(where k is the heat transfer coefficient, A is the heat dissipation area, T bolier = is the boiler surface temperature). It can be seen that the higher the ambient temperature, the smaller the heat loss (assuming that the boiler surface temperature is higher than the ambient temperature). Humidity has a certain impact on the combustion process in some biomass combustion situations. High humidity environment reduces the combustion efficiency of biomass fuel. These data are obtained through meteorological sensors installed near the boiler room so that the impact of environmental factors on boiler heat energy can be considered in subsequent analysis.

[0089] In the above S100, S130 of the data collection phase, the historical data collection obtains the operating data of the past period from the boiler control system or data recording system, including the historical records of the above parameters under different working conditions, to provide a data basis for subsequent model training and adaptive adjustment. Assume that the inlet water temperature at the i-th moment in the historical data is T in (i), the outlet water temperature is Tout (i), water flow rate is V(i), etc. These data are organized into time series data for analyzing the thermal energy variation law of the boiler under different operating conditions. By analyzing the parameter changes under different loads in the historical data, the characteristic patterns related to thermal energy output are found.

[0090] The S100, S140 in the data collection phase, outlier detection and correction: use a density-based local outlier factor (LOF) algorithm to detect outliers. For each data point, calculate the average value of the ratio of its local reachable density to the local reachable density of the neighborhood points (i.e., the LOF value). If the LOF value is much greater than 1, the point is determined to be an outlier. For the detected outliers, corrections are made in combination with the physical laws of boiler operation. If the inlet water temperature at a certain moment is abnormally high and exceeds the normal operating range, but according to the inlet water temperature change trend at the previous and subsequent moments and the relationship with other relevant parameters (such as fuel supply, steam pressure, etc.), it is judged that the abnormal spike is caused by a short-term sensor failure. The weighted average of adjacent normal data points can be used for correction. The weight is determined according to the time distance of the data points. The closer the distance, the greater the weight. Let the adjacent normal data points be T in (i-1) and T in (i+1), the time weights are w1 and w2 (satisfying w1+w2=1 and w1>w2), then the corrected inlet water temperature

[0091]

[0092] As a preferred solution of the present invention, S200, the model selection and training stage includes the following steps:

[0093] S210, hybrid model construction;

[0094] S220, multi-model fusion strategy;

[0095] S230, a method for determining a model fusion weight;

[0096] In the S200, model selection and training phase, S210, hybrid model construction: In the boiler thermal energy system, some factors have a linear relationship with the thermal energy output, while some have complex nonlinear characteristics. Under stable operation, there is a relatively direct linear correlation between water flow and thermal energy output, and the relationship is expressed as: linear =β0+β1x1+…+β n x n

[0097] Among them, y linear represents the thermal energy predicted by the linear part, x i Represents linearly related input parameters (such as water flow during stable operation, etc.) β iis the corresponding coefficient. Coefficient β i It can be estimated by the least squares method, and its calculation formula is:

[0098] β=(X T X) -1 X T y

[0099] Among them, X is the input parameter matrix (each row represents a set of input parameter samples, and each column corresponds to an input parameter), and y is the actual thermal energy output vector.

[0100] The fuel combustion process involves complex chemical reactions. There is a highly nonlinear relationship between the fuel composition, fuel efficiency and other factors and the heat release. Neural networks have advantages in dealing with such nonlinear problems. They construct a multi-layer perceptron structure, including an input layer, multiple hidden layers and an output layer. Suppose the input layer has n neurons (corresponding to n input parameters) and the jth hidden layer has m neurons. j neurons, and the output layer has 1 neuron (for predicting thermal energy). For the i-th neuron in the l-th layer (l=2,…,L, L is the total number of layers), its input With output The relationship is as follows:

[0101]

[0102] in, is the connection weight from the jth neuron in the (l-1)th layer to the ith neuron in the lth layer, is the bias of the ith neuron in the lth layer, and f(·) is the activation function (such as the ReLU function: f(x) = max(0,x)).

[0103] The final hybrid model prediction result is y = y linear +y neural In order to optimize the fusion effect, the weights of the two models can be dynamically adjusted according to their performance under different working conditions. Suppose the performance evaluation value of the linear model under the current working condition is p linear (It can be calculated based on mean square error, accuracy and other indicators). The performance evaluation value of the neural network model under the current working conditions is p neural , then the weight w of the linear model linear , and the weights w of the neural network model neural It can be calculated by the following formula:

[0104]

[0105] Under different operating conditions, the model weights can be automatically adjusted according to the performance, making the hybrid model better suited to different stages of boiler operation.

[0106] The linear regression model selected in the S200, model selection and training phase, and the multi-model fusion strategy is more accurate in describing the linear relationship and can provide a relatively stable prediction benchmark. The model formula is:

[0107] y LR =β0+β1x1+…+β n x n (Similar to the linear part in the mixed model above)

[0108] The support vector machine model has good generalization ability when dealing with small samples and nonlinear problems. Its prediction formula is: Where N is the number of support vectors, α i is the Lagrange multiplier, y i is the sample label, K(xi,x) is the kernel function (such as the radial basis kernel function K(xi,x) = exp(-γxi-x 2 ), γ is the kernel function parameter). In boiler thermal energy prediction, when the data samples are limited and there are complex nonlinear relationships, support vector machines can classify or regress data by finding the optimal hyperplane, and can achieve better results for data sets with complex boundaries.

[0109] The decision tree model has strong interpretability, can intuitively display the decision logic between input parameters and thermal energy output, and can process various types of data (including discrete and continuous data). Its prediction process is based on the gradual division of data features and makes decisions by building a tree structure. For fuel type (discrete data) and continuous data such as temperature and pressure, the decision tree can branch according to different feature thresholds and finally get the prediction results. This is very helpful for understanding the key factors and decision-making process in the boiler heat generation process, and it can also show certain advantages when dealing with some local nonlinear relationships.

[0110] Model selection

[0111] In the S200, S23 ... i Then, the stability index is included in the weight calculation. A simple way is to combine the stability index with the performance index (such as the mean square error e i ) to form a comprehensive evaluation index r i , such as r i =ei +λs i (λ is a parameter that weighs the importance of stability and performance). Determine the weight based on comprehensive evaluation indicators This allows the fusion model to focus more on model stability while pursuing accuracy, reducing the adverse impact on the overall prediction caused by large fluctuations in the performance of individual models under specific circumstances.

[0112] Further consider the adaptability of the model to different types of data patterns. Analyze the performance of each model when processing different feature combinations (such as data features under high temperature and high pressure conditions, data features when the fuel composition changes greatly, etc.). Give higher weights to models that perform well under specific data patterns. Historical data can be classified by cluster analysis and other methods to form different data patterns, and then the average performance indicators of each model on each cluster (such as accuracy, recall, etc.) are calculated, and the weights of the models under different data patterns are adjusted according to these indicators. When the boiler is under high load and unstable fuel quality conditions, if a model performs well under similar conditions in historical data, the weight of the model is appropriately increased under the current conditions, so that the fusion model can flexibly adjust the contribution of each model according to the actual operation data mode, and improve the pertinence and accuracy of the prediction. As time goes by and the boiler operation data accumulates, the model will experience performance drift. In order to capture and adapt to this change in a timely manner, a dynamic weight update strategy is adopted. Periodically (such as every period of time or after accumulating a certain number of new data points) re-evaluate the performance of each model on the latest data and update the weight. At the same time, in order to avoid model instability caused by too frequent weight updates, a forget factor μ (0 < μ < 1) can be introduced. i (t) is the model M i The weight at time t, e i (t) is the model M i The performance indicator (such as mean square error) calculated based on the latest data at time t, the weight update formula is The role of the forgetting factor is to gradually reduce the impact of historical data on the weight, so that the model weight can adapt to the new data distribution and performance changes more quickly, ensuring that the fusion model always maintains good prediction capabilities. In addition, considering the sudden abnormalities in the operation of the boiler, an anomaly detection mechanism is introduced to adjust the model weight. When an abnormal pattern in the data is detected (such as abnormal data points caused by sensor failure or boiler operation parameters suddenly deviate significantly from the normal range), the weight of the model that is more affected by the abnormality is reduced, while the weight of the model that is more robust to the abnormality is increased. A distance-based anomaly detection method (such as the K-nearest neighbor algorithm) is used to identify abnormal data points, and the prediction error distribution of each model in the area around the abnormal data point is calculated. If a model has a significantly larger error in the abnormal area, its weight is reduced, otherwise the weight is increased. In this way, when encountering abnormal situations, the fusion model can more reliably predict thermal energy and ensure the safety and stability of boiler operation.

[0113] Finally, in order to improve the adaptability and flexibility of the weight determination method, a reinforcement learning method is used. The weight adjustment is regarded as a decision-making process. The model learns the optimal weight allocation strategy based on the current boiler operating state (as the state space) and the predicted performance feedback (as the reward signal). Using the Q-learning algorithm, the state space is defined as the combination of the current key operating parameters of the boiler (such as temperature, pressure, fuel flow, etc.), and the action space is the adjustment operation of each model weight. By constantly trying different weight adjustment strategies and observing the changes in predicted performance (as rewards), the model gradually learns the optimal weight allocation method under different operating conditions, thereby achieving a more accurate prediction of boiler thermal energy and being able to automatically adapt to different operating conditions and data feature changes.

[0114] As a preferred solution of the present invention, the adaptive adjustment stage S300 includes the following steps:

[0115] S310, real-time data monitoring and working condition identification;

[0116] S320, dynamic adjustment of model parameters;

[0117] S330, adaptive optimization of model structure;

[0118] In the S300, S310 of the adaptive adjustment stage, the real-time data monitoring and working condition identification adopts the working condition identification based on wavelet transform and hidden Markov model (HMM): the real-time data of boiler operation (such as temperature, pressure, flow, etc.) is subjected to wavelet transform. Assume that the real-time data sequence is x(n), n=1,2,…,N, and the wavelet transform formula is Where a is the scale factor, b is the translation factor, and ψ(·) is the wavelet basis function. The data is decomposed into sub-bands of different frequencies through wavelet transform, and the energy characteristics E of the sub-bands are extracted.i ,i=1,2,…,k (k is the number of subbands). These energy features constitute the feature vector E=[E1,E2,…,E k ] T .

[0119] These feature vectors are used as the observation sequence of the Hidden Markov Model (HMM). The HMM consists of hidden states S = {s1, s2, ..., s M} (M is the number of states, such as startup, stable operation, variable load and other working conditions), observation symbol O = {o1, o2, ..., o L} (here the observation symbol can correspond to different energy characteristic value ranges), initial state probability distribution π = (π1, π2, ... π n )(π i =P(q1=s i ), q1 is the initial state), state transition probability matrix

[0120] A=(a ij )(a ij =P(q t+1 =s j q t =s j ),t=1,2,…) and the observation probability matrix B=(b ij )(b ij =P(o t =o j q t =s j ))composition.

[0121] The Baum-Welch algorithm is used to train HMM. The algorithm is based on maximum likelihood estimation and optimizes model parameters through iterative calculation to maximize the probability of the observation sequence P(Oλ) (λ=(A, B, π) is the HMMs parameter). After the training is completed, the Viterbi algorithm is used to calculate the most appropriate hidden state sequence for the new real-time data feature vector, thereby identifying the current operating conditions of the boiler and providing a basis for subsequent model adjustment.

[0122] The Adadelta optimization algorithm is used in the dynamic adjustment of model parameters in the S300 and S320 of the adaptive adjustment stage: During the model training process, the Adadelta optimization algorithm dynamically adjusts the learning rate by accumulating the square of the past gradient. Assume that the gradient of parameter θ at time t is g t , first calculate the exponential moving average of the square of the gradient (ρ is the reduction rate, usually between 0 and 1, such as 0.95).

[0123] Then calculate the parameter update amount Where η is the initial learning rate and ε is a minimum value (such as 10 -6 ) to prevent the denominator from being zero, E[θ 2 ] t-1 It is also an exponential moving average, and its update formula is The Adadelta optimization algorithm can adaptively adjust the learning rate based on the historical information of the gradient. When processing noisy and dynamically changing data such as boiler operation data, it can make the model parameter update more stable and effective, and improve the model's adaptability to new data.

[0124] The S300, adaptive adjustment stage, S330, model structure adaptive optimization adopts the neural network structure adjustment based on particle swarm optimization (PSO): for the neural network model structure (such as the number of hidden layers, the number of neurons in each layer, etc.), it is regarded as the particle position in the particle swarm optimization algorithm. Let the neural network structure parameter vector be X = [x1, x2, ... x D ] T (D is the structural parameter dimension, such as: x1 is the number of hidden layers, x2 is the number of neurons in the first layer, etc.).

[0125] Define the particle's velocity vector V = [v1, v2, ... v D ] T Each particle in the particle swarm updates its speed and position according to its own historical optimal position pbest and the global optimal position gbest of the swarm. The speed update formula is v i,d (t+1)=wv i,d (t)+c1r1(pbest i,d -x i,d (t))+c2r2(gbest d -x i,d (t)), where w is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers between [0,1], and t is the number of iterations. The position update formula is x i,d (t+1)=v i,d (t)+v i,d (t+1).

[0126] Substitute the neural network structure parameters into the model, calculate the model's performance indicators (such as mean square error MSE) on the validation set, and use the performance indicators as the particle's fitness function. Through multiple generations, the particle swarm continuously searches for the optimal neural network structure, enabling the model to automatically adjust its structure according to changes in boiler operating conditions, thereby improving the model's prediction performance and adaptability.

[0127] As a preferred solution of the present invention, in the prediction and evaluation stage S400, after the adaptive adjustment stage (S300), the model has optimized its own structure and parameters according to the real-time monitoring data and the working condition identification results, laying a solid foundation for accurate thermal energy output prediction. It includes the following steps:

[0128] S410, heat output prediction;

[0129] S420, performance evaluation index calculation;

[0130] S430, Result visualization and analysis;

[0131] In the S400, prediction and evaluation phase, S410, thermal output prediction, a prediction model fusion is performed in combination with working condition adaptation, and based on the working conditions identified in S400, a multi-model fusion strategy is further optimized for thermal output prediction. For different working conditions, different initial weights are assigned to each basic model (such as a neural network model, a support vector regression model, etc.). Under stable operating conditions, historical data shows that the neural network model performs well, so it is given a higher initial weight; and when the load changes drastically, it has been verified that the support vector regression model is more adaptable, so its weight is increased accordingly. Assume w i,k is the initial weight of model i under working condition k, which is determined by normalizing the inverse of the model performance evaluation index (such as mean square error) under the same historical working condition. As new data continues to flow in, the prediction performance of each model under the current working condition is updated in real time. The sliding window technology is used to calculate the performance index of the most recent p time steps (p can be set according to the dynamic characteristics of the data) to dynamically adjust the model weight. Assume that the predicted value of model i at time t under working condition k is The fusion prediction value (m is the total number of models). This prediction model fusion method combined with operating condition adaptation can fully utilize the advantages of different models under different operating conditions, improve the accuracy and stability of prediction, and effectively cope with the complex changes in boiler operation.

[0132] Data enhancement is performed using a large amount of different working condition data accumulated by real-time data monitoring in S400. For some key parameters (such as temperature, pressure, etc.), new sample data is generated by random perturbation, adding noise, etc., while ensuring that the new data is still within a reasonable physical range. Then, the transfer learning technology is used to migrate the model parameters trained under the working condition with rich historical data to the current new working condition or working condition model with less data as the initial parameters. Let the source working condition model parameter be θ s , the target operating condition model parameter is θ t, by calculating the similarity sim between the source working condition and the target working condition data (which can be calculated based on data distribution characteristics or feature correlation), the target working condition model parameters are updated by weighted average: (α is a weight parameter, which is adjusted according to the actual situation). On this basis, the model is fine-tuned and trained using the enhanced target working condition data to obtain a prediction model suitable for the current working condition, thereby improving the prediction ability under different working conditions, especially new working conditions or data-scarce working conditions.

[0133] The dynamic evaluation indicators of the changes in working conditions are considered in the calculation of the performance evaluation indicators in S400, S420 of the prediction and evaluation stage. In order to more accurately evaluate the performance of the model under different working conditions, dynamic evaluation indicators are designed. In addition to the traditional mean square error (MSE), the influencing factors of working condition changes are introduced. The change rate of the model prediction error Δe when the working condition switches or fluctuates is calculated. k (k represents the operating condition change point), as well as the mean and standard deviation of the prediction error during the stable operating condition period". Comprehensive performance evaluation indicators (β, γ, and δ are weight coefficients, which are determined according to the importance of each part under different working conditions.) This indicator can not only measure the overall prediction accuracy of the model, but also reflect the model's adaptability and stability to changes in working conditions, providing a more targeted basis for model optimization.

[0134] The model prediction value and the actual value are regarded as two probability distributions, and the degree of probability distribution matching between the two is calculated as the performance evaluation index. Metrics such as KL divergence (Kullback-Leibler Divergence) or Wasserstein distance are used. Suppose the probability distribution of the predicted value is P y , the probability distribution of the actual value is Q y , taking KL divergence as an example, the evaluation index The smaller the index is, the closer the probability distribution of the predicted value is to the actual value, and the better the model performance is. This evaluation method based on probability distribution matching can evaluate the model performance from the perspective of overall distribution, avoiding the limitations of focusing only on single-point errors, and is particularly suitable for evaluating boiler heat output prediction results with uncertainty or large fluctuations.

[0135] The S400, S430, results visualization and analysis in the prediction and evaluation phase create a visualization interface to closely link the prediction results with the changes in operating conditions. The predicted thermal output curve, the actual measurement value curve, and the operating condition switching mark point are displayed synchronously on the timeline, and different colors or line types are used to distinguish the predicted and actual data segments under different operating conditions, so as to facilitate intuitive observation of the performance differences of the model under various operating conditions. At the same time, the dynamic changes of operating condition-related parameters (such as load change curves, fuel composition changes, etc.) are displayed in the interface. Through interactive operations (such as clicking on the operating condition switching point), detailed parameter comparison information and predicted performance indicator changes are popped up to help users deeply understand the impact of operating condition changes on thermal output and the adaptability of the model.

[0136] The forecast error is decomposed into components, including trend error, periodic error, and random error (which can be obtained through time series decomposition methods). The changes of each error component over time are displayed in the form of stacked bar charts or line charts in the visualization interface. For trend error, analyze its relationship with the long-term operation trend of the boiler (such as efficiency changes caused by equipment aging); for periodic error, explore its correlation with the boiler operation cycle (such as daily or weekly load cyclical changes). Through this visualization method, users can clearly identify the main sources of error, so as to optimize the model or boiler operation in a targeted manner.

[0137] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A prediction method based on adaptive boiler thermal energy, comprising the following steps: S100, data collection and preprocessing stage; S200, model selection and training phase; S300, adaptive adjustment stage; S400, prediction and evaluation phase.

2. The prediction method according to claim 1, characterized in that: The data collection phase S100 includes the following steps: S110, parameter collection; S120, environmental data collection; S130, historical data collection; S140, outlier detection and correction; The basic parameters of the boiler in the S100, S110, parameter collection in the data collection phase include boiler model, capacity, design thermal efficiency, heating area, etc. These parameters determine the basic performance and potential heat energy generation capacity of the boiler. Fuel-related data: including fuel type (such as coal, natural gas, biomass, etc.), fuel calorific value (the amount of heat released by complete combustion of fuel per unit mass or volume), and fuel supply rate. Fuel calorific value is the key to calculating heat input. Assume that the fuel calorific value is q (k / kg or k / m 3 ), the fuel supply rate is m (kg / s or m 3 / s), then the heat provided by the fuel is Q fuel =q×m. Operation parameters: record the boiler inlet water temperature T in , water outlet temperature T out . Water flow rate V, pressure (steam pressure or hot water pressure) P, burner power (gas or oil boiler) W, etc. The inlet water temperature and water flow rate determine the amount of water to be heated and the initial state. The outlet water temperature reflects the heat transfer effect. According to the heat calculation formula Q = c × m × ΔT (where c is the specific heat capacity of water, m is the mass of water, and ΔT is the temperature change), the heat absorbed or released by water can be calculated, which is then associated with the boiler heat energy. S100, S120 of the data collection phase, and the ambient temperature T during the environmental data collection env , and humidity H env Data: Ambient temperature affects boiler heat dissipation loss. According to the heat conduction formula Q loss = k × A × (T bolier -T env )(where k is the heat transfer coefficient, A is the heat dissipation area, T bolier = is the boiler surface temperature). It can be seen that the higher the ambient temperature, the smaller the heat loss (assuming that the boiler surface temperature is higher than the ambient temperature). Humidity has a certain impact on the combustion process in some biomass combustion situations. High humidity environment reduces the combustion efficiency of biomass fuel. These data are obtained through meteorological sensors installed near the boiler room so that the impact of environmental factors on boiler heat energy can be considered in subsequent analysis. In the above S100, S130 of the data collection phase, the historical data collection obtains the operating data of the past period from the boiler control system or data recording system, including the historical records of the above parameters under different working conditions, to provide a data basis for subsequent model training and adaptive adjustment. Assume that the inlet water temperature at the i-th moment in the historical data is T in (i), the outlet water temperature is T out (i), water flow rate is V(i), etc. These data are organized into time series data for analyzing the thermal energy variation law of the boiler under different operating conditions. By analyzing the parameter changes under different loads in the historical data, the characteristic patterns related to thermal energy output are found. S100, S140 of the data collection phase, outlier detection and correction: use the density-based local outlier factor (LOF) algorithm to detect outliers. Let the adjacent normal data points be T in (i-1) and T in (i+1), the time weights are w1 and w2 (satisfying w1+w2=1 and w1>w2), then the corrected inlet water temperature 3. The prediction method according to claim 1, characterized in that: The S200, model selection and training phase includes the following steps: S210, hybrid model construction; S220, multi-model fusion strategy; S230, a method for determining a model fusion weight; In the S200, model selection and training phase, S210, hybrid model construction: In the boiler thermal energy system, some factors have a linear relationship with the thermal energy output, while some have complex nonlinear characteristics. Under stable operation, there is a relatively direct linear correlation between water flow and thermal energy output, and the relationship is expressed as: linear =β0+β1x1+…+β n x n Among them, y linear represents the thermal energy predicted by the linear part, x i Represents linearly related input parameters (such as water flow during stable operation, etc.) β i is the corresponding coefficient. Coefficient β i It can be estimated by the least squares method, and its calculation formula is: β=(X T X) -1 X T y Among them, X is the input parameter matrix (each row represents a set of input parameter samples, and each column corresponds to an input parameter), and y is the actual thermal energy output vector. Assume that the input layer has n neurons (corresponding to n input parameters) and the jth hidden layer has m j neurons, and the output layer has 1 neuron (for predicting thermal energy). For the i-th neuron in the l-th layer (l=2,…,L, L is the total number of layers), its input With output The relationship is as follows: in, is the connection weight from the jth neuron in the (l-1)th layer to the ith neuron in the lth layer, is the bias of the ith neuron in the lth layer, and f(·) is the activation function (such as the ReLU function: f(x) = max(0,x)). The final hybrid model prediction result is y = y linear +y neural In order to optimize the fusion effect, the weights of the two models can be dynamically adjusted according to their performance under different working conditions. Suppose the performance evaluation value of the linear model under the current working condition is p linear (It can be calculated based on mean square error, accuracy and other indicators). The performance evaluation value of the neural network model under the current working conditions is p neural , then the weight w of the linear model linear , and the weights w of the neural network model neural It can be calculated by the following formula: Under different operating conditions, the model weights can be automatically adjusted according to the performance, making the hybrid model better suited to different stages of boiler operation. The linear regression model selected in the S200, model selection and training phase, and the multi-model fusion strategy is more accurate in describing the linear relationship and can provide a relatively stable prediction benchmark. The model formula is y LR =β0+β1x1+…+β n x n , the support vector machine model has good generalization ability when dealing with small samples and nonlinear problems. Its prediction formula is Where N is the number of support vectors, α i is the Lagrange multiplier, yi is the sample label, K(x i ,x) is a kernel function (such as the radial basis kernel function K(x i ,x)=exp(-γ||x i -x|| 2 ), γ is the kernel function parameter). In boiler thermal energy prediction, when the data samples are limited and there are complex nonlinear relationships, support vector machines can classify or regress data by finding the optimal hyperplane, and can achieve better results for data sets with complex boundaries. S200, S230 in the model selection and training phase, and the method for determining the weight of the model fusion weight based on the comprehensive evaluation index Assume w i (t) is the model M i The weight at time t, e i (t) is the model M i The performance indicator (such as mean square error) calculated based on the latest data at time t, the weight update formula is 4. The prediction method according to claim 1, characterized in that: The adaptive adjustment stage S300 includes the following steps: S310, real-time data monitoring and working condition identification; S320, dynamic adjustment of model parameters; S330, adaptive optimization of model structure; In the S300, S310 of the adaptive adjustment stage, the real-time data monitoring and working condition identification adopts the working condition identification based on wavelet transform and hidden Markov model (HMM): the real-time data of boiler operation (such as temperature, pressure, flow, etc.) is subjected to wavelet transform. Assume that the real-time data sequence is x(n), n=1,2,…,N, and the wavelet transform formula is Where a is the scale factor, b is the translation factor, and ψ(·) is the wavelet basis function. The data is decomposed into sub-bands of different frequencies through wavelet transform, and the energy characteristics E of the sub-bands are extracted. i ,i=1,2,…,k (k is the number of subbands). These energy features constitute the feature vector E=[E1,E2,…,E k ] T . These feature vectors are used as the observation sequence of the Hidden Markov Model (HMM). The HMM consists of hidden states S = {s1, s2, ..., s M } (M is the number of states, such as startup, stable operation, variable load and other working conditions), observation symbol O = {o1, o2, ..., o L } (here the observation symbol can correspond to different energy characteristic value ranges), initial state probability distribution π = (π1, π2, ... π n )(π i =P(q1=s i ), q1 is the initial state), state transition probability matrix A = (a ij )(a ij =P(q t+1 =s j |q t =s j ),t=1,2,…) and the observation probability matrix B=(b ij )(b ij =P(o t =o j |q t =s j ))composition. The Adadelta optimization algorithm is used in the dynamic adjustment of model parameters in the S300 and S320 of the adaptive adjustment stage: During the model training process, the Adadelta optimization algorithm dynamically adjusts the learning rate by accumulating the square of the past gradient. Assume that the gradient of parameter θ at time t is g t , first calculate the exponential moving average of the square of the gradient (ρ is the reduction rate, usually between 0 and 1, such as 0.95). Then calculate the parameter update amount Where η is the initial learning rate and ε is a minimum value (such as 10 -6 ) to prevent the denominator from being zero, E[θ 2 ] t-1 It is also an exponential moving average, and its update formula is The Adadelta optimization algorithm can adaptively adjust the learning rate based on the historical information of the gradient. When processing noisy and dynamically changing data such as boiler operation data, it can make the model parameter update more stable and effective, and improve the model's adaptability to new data. The S300, S330, model structure adaptive optimization in the adaptive adjustment stage adopts a neural network structure adjustment based on particle swarm optimization (PSO): let the neural network structure parameter vector be X=[x1, x2, ...x D ] T (D is the structural parameter dimension, such as: x1 is the number of hidden layers, x2 is the number of neurons in the first layer, etc.). Define the particle's velocity vector V = [v1, v2, ... v D ] T Each particle in the particle swarm updates its speed and position according to its own historical optimal position pbest and the global optimal position gbest of the swarm. The speed update formula is v i,d (t+1)=wv i,d (t)+c1r1(pbest i,d -x i,d (t))+c2r2(gbest d -x i,d (t)), where w is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers between [0,1], and t is the number of iterations. The position update formula is x i,d (t+1)=v i,d (t)+v i,d (t+1). Substitute the neural network structure parameters into the model, calculate the performance index of the model on the validation set, and use the performance index as the fitness function of the particle. Through multiple generations, the particle swarm continuously searches for the optimal neural network structure, so that the model can automatically adjust the structure according to the changes in the boiler operating conditions, thereby improving the prediction performance and adaptability of the model.

5. The prediction method according to claim 1, characterized in that: In the prediction and evaluation phase S400, after the adaptive adjustment phase (S300), the model has optimized its own structure and parameters based on real-time monitoring data and working condition identification results, laying a solid foundation for accurate thermal output prediction. It includes the following steps: S410, heat output prediction; S420, performance evaluation index calculation; S430, Result visualization and analysis; In the S400, prediction and evaluation phase, S410, heat output prediction, a prediction model fusion combining working condition adaptation is performed, and w i,k is the initial weight of model i under working condition k, which is determined by normalizing the inverse of the model performance evaluation index (such as mean square error) under the same historical working condition. As new data continues to flow in, the prediction performance of each model under the current working condition is updated in real time. The sliding window technology is used to calculate the performance index of the most recent p time steps (p can be set according to the dynamic characteristics of the data) to dynamically adjust the model weight. Assume that the predicted value of model i at time t under working condition k is The fusion prediction value (m is the total number of models). Assume that the source condition model parameter is θ s , the target operating condition model parameter is θ t , by calculating the similarity sim between the source working condition and the target working condition data (which can be calculated based on data distribution characteristics or feature correlation), the target working condition model parameters are updated by weighted average: (α is a weight parameter, which is adjusted according to the actual situation). On this basis, the model is fine-tuned and trained using the enhanced target working condition data to obtain a prediction model suitable for the current working condition, thereby improving the prediction ability under different working conditions, especially new working conditions or data-scarce working conditions. The S400, S420 of the prediction and evaluation phase, the dynamic evaluation index considering the change of working conditions in the performance evaluation index calculation, the comprehensive performance evaluation index (β, γ, δ are weight coefficients, which are determined according to the importance of the impact of each part under different working conditions). The degree of probability distribution matching between the two is calculated as the performance evaluation index. Assume that the probability distribution of the predicted value is P y , the probability distribution of the actual value is Q y , taking KL divergence as an example, the evaluation index The smaller the index is, the closer the probability distribution of the predicted value is to the actual value, and the better the model performance is. The S430 results visualization and analysis of the S400 prediction and evaluation phase: Create a visualization interface to associate the prediction results with the operating condition changes, and display the heat output prediction, actual measurement value curve and operating condition switching mark points on the time axis simultaneously. Use color or line type to distinguish different operating condition data segments, display the dynamic changes of operating condition related parameters, and interactive operations can pop up detailed information. Decompose the prediction error components (using the time series decomposition method), and display the error components over time in the interface with stacked bar charts or line charts, analyze the relationship between trend errors and long-term boiler operation trends, and the correlation between period errors and operating cycles, to help users identify error sources and optimize model or boiler operation.

Images