Intelligent load prediction, fault diagnosis and optimal control method for heat supply system

Abstract

The invention relates to the technical field of heat supply system operation management, in particular to an intelligent load prediction, fault diagnosis and optimal control method for a heat supply system, which comprises the following steps of: 1, acquiring multi-dimensional data; 2, data preprocessing; step 3, load prediction: based on the preprocessed multi-dimensional time sequence data set output in the step 2, a long short-term memory network model is used to predict a thermal load prediction value sequence in a future time period, and the input of the long short-term memory network model comprises historical operation parameters, historical environment parameters and time characteristic data; step 4, fault diagnosis; and 5, optimization control is conducted, specifically, a control instruction is generated through a model prediction control method, a dynamic model of the heat supply system is built through the model prediction control method, control variable set values are optimized, and control variables comprise the pump rotating speed, the valve opening degree and the heating power. By integrating data acquisition, prediction, diagnosis and optimization control, the heat supply system can operate efficiently and stably.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of operation management of heating systems, and particularly to an intelligent load prediction, fault diagnosis and optimal control method for heating systems. BACKGROUND

[0002] As an important part of urban infrastructure, heating systems bear the key task of providing stable heat energy to users. Traditional heating systems usually use simple statistical methods based on historical data for load prediction, such as using linear regression or time series analysis. These methods rely on heating data from the past few years, but ignore the dynamic impact of real-time environmental factors, user behavior changes and equipment operating status. Therefore, the prediction accuracy of traditional load prediction methods significantly decreases when dealing with sudden weather changes, holiday load fluctuations or equipment aging, leading to insufficient heating or energy waste. In terms of fault diagnosis, existing technologies mainly rely on manual inspection and regular maintenance, collecting basic parameters such as temperature and pressure through sensors, and determining fault types based on the experience of technicians. This method is inefficient and cannot detect potential faults in real time, such as pipe leaks or pump valve failures, which are often discovered after the fault occurs, causing heating interruptions and increased maintenance costs. In terms of optimal control, traditional control strategies are based on fixed set values or simple feedback control, which cannot dynamically adjust operating parameters according to predicted loads and fault risks, resulting in low system energy efficiency and high operating costs. For example, when the load is low, the system still runs at high power, wasting energy; or in the early stages of a fault, the control strategy fails to adjust in time, exacerbating equipment damage.

[0003] In existing technologies, some improved methods attempt to introduce machine learning algorithms for load prediction, but these methods usually handle prediction, diagnosis and control problems in isolation, lacking integration. For example, a load prediction model may use a neural network, but it is not linked to a fault diagnosis module, resulting in prediction results that cannot be used for real-time fault detection. In addition, these methods have limited data collection range, often focusing only on temperature and historical load, ignoring multi-dimensional parameters such as flow, pressure and equipment vibration, resulting in insufficient diagnostic accuracy. At the same time, optimal control strategies are often based on rule bases, lacking adaptive learning ability, and cannot cope with system aging or external environmental changes. Overall, existing technologies have the following core problems: first, low load prediction accuracy, due to insufficient data dimensions and model staticity, which cannot adapt to dynamic environments; second, delayed fault diagnosis, due to reliance on manual and simple threshold detection, which cannot achieve early warning; third, rigid control strategies, due to lack of linkage with prediction and diagnosis results, which cannot achieve energy efficiency optimization. These problems result in low operating efficiency, high failure rate and serious energy waste of heating systems, necessitating an integrated intelligent method to address them. SUMMARY

[0004] In order to achieve the above purposes, the present application provides an intelligent load prediction, fault diagnosis and optimal control method for a heating system, comprising the following steps: Step 1: Multi-dimensional data acquisition, collecting real-time operating parameters and environmental parameters of the heating system, including temperature data, flow data, pressure data, vibration data, energy consumption data and environmental data, all data being collected by sensors and data acquisition devices at a preset sampling frequency and transmitted to a data processing unit to generate a multi-dimensional time series data set; Step 2: Data preprocessing, data cleaning, data filling, data standardization and data alignment are performed on the multi-dimensional time series data set output by step 1, abnormal values are identified and removed, missing data is supplemented using an interpolation method, data is scaled to the same numerical range, and data alignment ensures that the time stamps of all parameters are consistent, generating a preprocessed multi-dimensional time series data set; Step 3: Load prediction, based on the preprocessed multi-dimensional time series data set output by step 2, a long short-term memory network model is used to predict the heat load prediction value sequence of the future period, the input of the long short-term memory network model including historical operating parameters, historical environmental parameters and time feature data; Step 4: Fault diagnosis, based on the preprocessed multi-dimensional time series data set output by step 2 and the heat load prediction value sequence output by step 3, features in the multi-dimensional time series data set are extracted and input into a support vector machine model, the support vector machine model outputting fault types and fault probabilities; Step 5: Optimal control, based on the heat load prediction value sequence output by step 3 and the fault diagnosis result output by step 4, a model predictive control method is used to generate control instructions, the model predictive control method constructing a dynamic model of the heating system and optimizing control variable set values, the control variables including pump speed, valve opening and heating power, and the optimization objectives including maximum energy efficiency and minimum fault risk.

[0005] Preferably, the specific process of multi-dimensional data acquisition in step 1 includes: temperature data is collected by temperature sensors to collect the inlet water temperature and return water temperature of the heating pipe network; flow data is collected by flow sensors to collect flow data of the heating pipe network; pressure data is collected by pressure sensors to collect pressure data of the heating pipe network; vibration data is collected by vibration sensors to collect vibration data of pumps and valves; energy consumption data is collected by energy consumption metering devices to collect heat energy consumption data; environmental data is collected by environmental sensors to collect outdoor temperature and humidity data; all sensor data is collected by data acquisition devices at a preset sampling frequency and transmitted to a data processing unit, time stamps are marked for each parameter during data acquisition to ensure data synchronization; the preset sampling frequency is dynamically adjusted according to the operating period of the heating system and external environmental changes, the dynamic adjustment process being based on historical data analysis and real-time load demand, and the optimal sampling interval being calculated by the data processing unit.

[0006] Preferably, the specific process of data preprocessing in step 2 includes: data cleaning uses a sliding window method to calculate the mean and standard deviation of each parameter, and marks and deletes data points exceeding three times the standard deviation as abnormal; data filling uses a linear interpolation method to supplement missing data, and the linear interpolation is based on the trend of adjacent data points; data standardization adopts a min-max normalization method to convert each parameter to a numerical range between zero and one; data alignment ensures the consistency of time series of all parameters through a timestamp matching algorithm, and the timestamp matching algorithm uses a nearest neighbor interpolation method to align data points with different sampling frequencies; the preprocessed multi-dimensional time series data set is stored in a database for subsequent steps.

[0007] Preferably, the specific process of load prediction in step 3 includes: the input of the long short-term memory network model includes historical temperature data, historical flow data, historical pressure data, historical energy consumption data, historical environmental data, and time feature data, including date type and holiday flag; the long short-term memory network model learns the nonlinear relationship in the data through a training process, the training process uses a historical multi-dimensional time series data set, and the model parameters are optimized through a back propagation algorithm, which minimizes the prediction error based on the gradient descent method; the model output is a sequence of future heat load prediction values for a specified period, and the heat load prediction value is expressed in unit time heat demand; the number of hidden layer nodes of the long short-term memory network model is dynamically determined according to the data dimension and the prediction accuracy requirement, and the dynamic determination process selects the optimal number of nodes through a cross-validation method.

[0008] Preferably, the specific process of fault diagnosis in step 4 includes: extracting features from the multi-dimensional time series data set, including temperature change rate, flow fluctuation amplitude, pressure gradient, vibration spectrum feature, and deviation of heat load prediction value from actual value; the support vector machine model learns the patterns of normal state and multiple fault states through a training process, and the training data includes historical fault records and corresponding feature data; the support vector machine model outputs fault type and fault probability, and the fault type includes pipe blockage, pump failure, valve leakage, and heat exchanger efficiency decline; the fault probability is calculated through a decision function of the support vector machine model, and the decision function is based on the distance of the feature vector and the classification hyperplane; the fault diagnosis result is updated in real time and compared with the historical fault database to improve the diagnosis accuracy.

[0009] Preferably, the specific process of optimizing control in step 5 includes: a model predictive control method constructs a dynamic model of the heating system, the dynamic model is based on the principles of mass conservation and energy conservation, the inputs include predicted values of heat load, fault diagnosis results and current operating parameters, and the output is the set value of the control variable; an optimization solver solves the set value of the control variable with the maximum energy efficiency and the minimum fault risk as the objective function, the maximum energy efficiency in the objective function is represented as the matching degree of actual heating and demand heat load, and the minimum fault risk is represented as the weighted sum of fault probability; the optimization solving process iteratively updates the set value of the control variable using the gradient descent method until the objective function converges; the set value of the control variable is sent to the actuator to adjust the pump speed, valve opening and heating power; the actuator feeds back the control result to the data processing unit to realize closed-loop control.

[0010] Preferably, the training process of the long short-term memory network model further includes: using an adaptive learning rate adjustment strategy, the adaptive learning rate is dynamically adjusted according to the change of training error, and the dynamic adjustment process is based on an exponential decay method; the input data of the long short-term memory network model generates training samples through a sliding window method, and the sliding window size is determined according to the inertia characteristics of the heating system and external environmental changes; model validation uses the leave-one-out method to divide the data set into a training set and a test set, and the test set is used to evaluate the model generalization ability; overfitting prevention is achieved by adding a regularization term to the loss function, and the regularization term is based on the L2 norm of the weight.

[0011] Preferably, the training process of the support vector machine model further includes: using a kernel function to process nonlinear features, the kernel function selects a radial basis function, and the parameters of the radial basis function are optimized by a grid search method; the classification hyperplane of the support vector machine model is determined by solving a convex optimization problem, and the convex optimization problem is solved using a sequential minimal optimization algorithm; the real-time of fault diagnosis is realized through parallel computing of feature extraction, and the parallel computing uses a multi-thread processing technology based on task decomposition and data partitioning; the output of the support vector machine model also includes fault severity, which is calculated by the length of the feature vector and the classification confidence.

[0012] Preferably, the dynamic model of the model predictive control method further includes: using a state space equation to represent the dynamic behavior of the heating system, the state space equation includes a mass balance equation and an energy balance equation; the objective function of the optimization solver further includes device life extension, which is realized by a smoothness constraint of the control variable, and the smoothness constraint is represented as a limit of the control variable change rate; the optimization solving process uses an interior point method to handle the constraint condition, and the interior point method converts the constraint problem into an unconstrained problem through an obstacle function; the generation frequency of the control command is dynamically adjusted according to the system response time and external disturbance, and the dynamic adjustment process is based on real-time data acquisition and prediction error feedback.

[0013] Preferably, the linkage process of step 3, step 4 and step 5 comprises: the load prediction result is input into the fault diagnosis step in real time, and the fault diagnosis result is input into the optimal control step in real time; the optimal control step adjusts the control strategy according to the fault probability, and when the fault probability exceeds a threshold value, the optimal control step preferentially executes a fault mitigation strategy; the fault mitigation strategy comprises reducing the pump rotating speed, adjusting the valve opening degree and reducing the heating power; the whole linkage process realizes real-time data stream processing through a data processing unit, the data stream processing uses a pipeline architecture, and the pipeline architecture ensures seamless data transmission between steps; system performance monitoring is realized through a log recording and alarm mechanism, and the log recording comprises prediction accuracy, diagnosis accuracy and control efficiency indicators.

[0014] Advantages of the present application: 1. The load prediction accuracy is improved, because the multi-dimensional data acquisition and the adaptive long short-term memory network model are adopted, the dynamic environmental changes and user behaviors can be accurately captured, and the prediction error is reduced.

[0015] 2. The timeliness of fault diagnosis is enhanced, because the support vector machine model can early identify potential faults by combining real-time data and prediction results, and reduce system downtime.

[0016] 3. The control strategy is optimized, because the model predictive control method can dynamically adjust the operating parameters by integrating the prediction and diagnosis results, improve the energy efficiency and reduce the fault risk. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the present application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0018] Fig. 1 The step flow chart of the method of the present application; Fig. 2 The step flow chart of the fault diagnosis in step 4 of the method of the present application. DETAILED DESCRIPTION

[0019] The present application will be described in detail below in combination with the drawings and specific embodiments. It should be noted here that in order to make the embodiments more detailed, the following embodiments are the best and preferred embodiments, and other alternative ways can also be used by those skilled in the art to implement them; and the drawings are only used to more specifically describe the embodiments, and are not intended to specifically limit the present application.

[0020] Please refer to Figs. 1-2The embodiments of the present application provide an intelligent load prediction, fault diagnosis and optimization control method for a heating system. In step 1, in the heating system, accurate real-time data collection is the basis for ensuring intelligent management of the system. By laying out sensors to collect various real-time operating parameters (such as temperature, flow, pressure, vibration, energy consumption) and environmental parameters (such as external temperature, humidity, etc.), these data are automatically collected and transmitted to the data processing unit according to the preset sampling frequency (for example, collected once every minute) through the data collection device. The collected data form a multi-dimensional time series data set, providing a basis for subsequent analysis and control.

[0021] Through the collection of multi-dimensional data, the working state of the heating system and environmental changes can be comprehensively reflected, providing high-quality input data to support subsequent steps. Comprehensive collection helps to achieve accurate load prediction and fault diagnosis, enhancing the reaction capability of the system.

[0022] In step 2, data preprocessing is a key link to ensure the accuracy of the analysis results. First, data cleaning is performed to eliminate abnormal values that may occur during the collection process (such as abnormal data caused by sensor failure). Second, interpolation methods are used to fill in missing data to ensure data integrity. Then, standardization processing is used to convert parameters of different dimensions to a unified numerical range, facilitating subsequent analysis. Finally, the data is aligned to ensure that all data timestamps are consistent, so that the time series data of different parameters can be correctly synchronized, ensuring the timeliness and accuracy of the subsequent analysis results.

[0023] Data preprocessing ensures the quality of the data, avoids errors caused by abnormal values and missing values, and improves the accuracy and reliability of model analysis.

[0024] In step 3, in the load prediction link, a long short-term memory (LSTM) model is used for time series prediction. LSTM network is a neural network suitable for long-time dependent data, which can predict future heat load demand based on historical operating parameters, environmental parameters and time feature data. By learning the rules in historical data, the LSTM model can accurately capture the changing trend of the heating system under different environmental and load conditions, thereby providing accurate load prediction results for subsequent optimization control.

[0025] The application of LSTM model can effectively improve the accuracy of load prediction, so that the heating system can make dynamic adjustments according to the accurate load demand, avoid energy waste or insufficient heating phenomenon, and improve the overall energy-saving effect of the system.

[0026] In step 4, fault diagnosis is performed by inputting the preprocessed data and load prediction results into a support vector machine (SVM) model. The SVM model, based on multi-dimensional feature extraction, can identify the type of fault in the heating system and assess the probability of failure. By learning from historical system fault data, the SVM can identify various fault patterns (such as equipment failure, pipe leakage, etc.) and provide timely fault warnings.

[0027] Through the fault diagnosis function, the system can achieve early warning, avoid equipment failure and downtime, and reduce maintenance costs and downtime. Compared with traditional manual detection methods, intelligent diagnosis can greatly improve detection efficiency and accuracy.

[0028] In step 5, in the optimization control step, a model predictive control (MPC) method is used to optimize the control strategy of the heating system based on load prediction and fault diagnosis results. The MPC method constructs a dynamic model of the heating system, which can adjust control variables (such as pump speed, valve opening, and heating power) according to real-time load demand, fault state, and system operating state. The optimization goal is to maximize energy efficiency and minimize fault risk, ensuring that the system operates efficiently and safely.

[0029] The MPC method adjusts system parameters in real time to keep the heating system running efficiently in dynamic environments, not only improving energy utilization efficiency, but also effectively reducing the risk of failure, further improving the economic and reliability of the system.

[0030] By combining multi-dimensional data collection, intelligent load prediction, fault diagnosis, and optimization control, an efficient and intelligent heating system management framework is established. The system can dynamically adjust the operation strategy based on real-time data and prediction results, improving energy utilization efficiency, reducing maintenance costs, and extending equipment life. At the same time, the intelligent level of the system improves the flexibility and adaptability of the overall operation, making the heating system more efficient, reliable, and energy-saving.

[0031] In one possible implementation, the inlet water temperature and return water temperature of the heating pipe network are important parameters for judging the operation efficiency of the heating system. By arranging temperature sensors, the inlet water and return water temperatures of the heating pipe network are monitored. These temperature data directly reflect the heat loss in the heat transfer process and can be used as an important basis for load prediction and optimization control.

[0032] Flow data is a key indicator of the amount of hot water flowing in the heating pipe network. By installing flow sensors, real-time flow data is collected. Flow data can reflect changes in system load demand and, combined with temperature data, can be used to further analyze the energy efficiency and load changes of the heating system.

[0033] Pressure variations in a heating system are closely related to the stability of the pipe network and the heating effect. Pressure sensors are used to monitor the pressure in the heating pipe network to ensure that the system operates within the normal pressure range. Pressure data can assist in fault diagnosis and help analyze whether there are problems such as pipe leaks or pump damage.

[0034] Vibration data acquisition is used to monitor the health status of critical equipment in heating systems, such as pumps and valves. Vibration sensors are installed at pumps and valves to collect vibration data in real time. Abnormal vibration data can provide early warnings of equipment failure, avoiding downtime and repair costs caused by equipment malfunctions.

[0035] Energy consumption data is monitored through energy metering devices to track the heat energy consumption of the entire heating system. This data helps assess system operating efficiency and provides a reference for optimized control. By monitoring heat energy consumption in real time, heating power can be precisely adjusted to achieve energy savings.

[0036] Collecting environmental data helps determine the impact of the external environment on the operation of the heating system. By collecting outdoor temperature and humidity data through environmental sensors, this data can reflect the impact of seasonal variations and weather conditions on heating demand, further optimizing load forecasting models.

[0037] All sensor data is collected by the data acquisition device at a preset sampling frequency and transmitted to the data processing unit in real time. During the acquisition process, each data point is timestamped to ensure accurate synchronization between different parameters. This synchronization plays a crucial role in the subsequent analysis, processing, and fault diagnosis of multi-dimensional data, avoiding errors caused by data asynchrony.

[0038] The preset sampling frequency is dynamically adjusted based on the heating system's operating cycle and changes in the external environment. For example, when heating demand is high or the external environment changes drastically, the system will increase the sampling frequency to acquire data more frequently; while during stable operation or when external environmental changes are minimal, the sampling frequency can be appropriately reduced to avoid excessive data collection and overburdening the system. This adjustment process relies on historical data analysis and real-time load demand, with the optimal sampling interval calculated by the data processing unit.

[0039] Through the efficient data acquisition and dynamic adjustment mechanism of multi-dimensional sensors, the entire heating system can achieve precise control and efficient operation under intelligent management, providing a sufficient data foundation for fault diagnosis and load forecasting.

[0040] In one possible implementation, data cleaning is the first step in the preprocessing stage, aimed at removing outliers and noisy data. To achieve this, a sliding window method is used. Specifically, for each acquired parameter, the mean and standard deviation are first calculated using a sliding window. Then, by judging the deviation of data points from the mean, data points deviating from the normal range by more than three times the standard deviation are marked. These data points are considered outliers and are deleted. The advantage of this method is that it can effectively remove unreasonable data caused by sensor malfunctions or environmental interference, ensuring the accuracy of subsequent data analysis.

[0041] During data acquisition, missing data points may occur, especially during long-term acquisition or in cases of sensor malfunction. To handle missing values, a linear interpolation method is employed. This method predicts the location of the missing value by calculating the trend of adjacent data points before and after it. Specifically, linear interpolation assumes that data changes are linear, thus filling in missing values ​​using the linear relationship between two consecutive data points. This method can smoothly supplement data and avoid analytical bias caused by missing data.

[0042] To ensure comparability of data across different parameters, data standardization is a necessary step. A min-max standardization method was employed, scaling the value of each parameter to a uniform range, from zero to one. This method allows data with different dimensions (such as temperature, pressure, and flow rate) to be compared and analyzed on the same scale, avoiding the impact of numerical range differences on subsequent modeling and analysis.

[0043] Because various sensors in a heating system may use different sampling frequencies, the timestamps of each parameter may be inconsistent. To address this issue, a timestamp matching algorithm is employed. This algorithm uses nearest-neighbor interpolation to ensure the alignment of data points with different sampling frequencies. Specifically, for data points with inconsistent timestamps, the algorithm calculates the closest sampling point based on the time difference, thereby synchronizing the time series of all parameters. This is crucial for subsequent multi-dimensional data analysis, as data from different time steps needs to be precisely matched to avoid misalignment or loss.

[0044] After cleaning, imputation, standardization, and alignment, the multi-dimensional time-series data is stored in a database for subsequent steps. This storage system ensures data integrity and efficient access, enabling subsequent load forecasting, fault diagnosis, and optimization control processes to be based on high-quality data.

[0045] The precise execution of the data preprocessing steps ensures the high quality and consistency of various sensor data in the heating system, thus providing a solid foundation for load forecasting, fault diagnosis, and optimized control.

[0046] In one possible implementation, the input data for the LSTM model of load forecasting includes data from multiple dimensions. First, historical temperature, flow rate, pressure, energy consumption, environmental data, and time-related data are all used as input. This data provides sufficient background information to help the model capture various factors influencing heat load demand. In particular, time-related data, such as date types and holiday markers, helps the model identify periodic and seasonal variations. For example, energy consumption patterns typically differ between holidays and specific dates, and the model can use these characteristics to more accurately predict fluctuations in heat load demand.

[0047] The LSTM model is a deep learning algorithm capable of effectively processing and predicting nonlinear time-series data. The model learns nonlinear relationships in historical data through a training process. Specifically, training uses historical multi-dimensional time-series datasets. By memorizing long-term dependencies, the LSTM model effectively captures the complex correlations between parameters such as temperature, flow rate, and pressure. Through the backpropagation algorithm, the LSTM model continuously adjusts its internal parameters, optimizing its predictive ability. The backpropagation algorithm is based on gradient descent, which minimizes prediction error, allowing the model to gradually improve its accuracy in predicting future heat loads during training.

[0048] A trained LSTM model can output a sequence of predicted heat load values ​​for a specified future time period, expressed as heat demand per unit time. For example, the model may predict heat load demand for the next few hours or days, accurate to the hour or minute. These predictions provide a basis for the operation and scheduling of the heating system, helping operators adjust heat source supply in advance, ensuring efficient system operation, and avoiding resource waste or insufficient heating.

[0049] The performance of an LSTM model largely depends on its structure, particularly the number of hidden layer nodes. To achieve optimal prediction accuracy, the number of hidden layer nodes is not fixed but dynamically determined based on the data dimensionality and required prediction accuracy. This dynamic determination is achieved through cross-validation. Cross-validation involves dividing the data into multiple subsets, repeatedly training and testing the model, and selecting the number of hidden layer nodes that provides the best prediction results. This method ensures that the LSTM model is not affected by overfitting or underfitting, thus improving the model's stability and generalization ability.

[0050] The application of LSTM models in load forecasting significantly improves the accuracy of forecasts and the operating efficiency of the system by effectively learning nonlinear patterns and long-term dependencies in historical data, laying the foundation for intelligent scheduling and optimized control of heating systems.

[0051] In one possible implementation, the fault diagnosis process first requires extracting several key features from a multi-dimensional time-series dataset. These features reflect anomalies in system operation and help the model identify fault types. Common features include: Temperature change rate: This indicates the rate at which the temperature of the heating system changes. Too fast or too slow a temperature change may be a signal of a system malfunction.

[0052] Flow fluctuation range: Instability in flow may indicate problems such as pipe blockage or pump failure.

[0053] Pressure gradient: Abnormal changes in pressure are usually a sign of pipeline leaks or decreased pump efficiency.

[0054] Vibration spectrum characteristics: By analyzing the vibration spectrum of equipment, faults in pumps, valves or other mechanical equipment can be detected, especially signs of wear or damage.

[0055] Deviation between predicted and actual heat load: If there is a significant difference between the actual and predicted heat load of the system, it may indicate a decrease in heat exchanger efficiency or other malfunctions.

[0056] These features provide rich information for subsequent fault diagnosis, helping the support vector machine (SVM) model to identify different fault modes.

[0057] Support Vector Machines (SVMs) are a machine learning method commonly used for classification problems. In fault diagnosis, SVM models are trained by learning feature patterns from normal states and various fault states. Training data includes historical fault records and their corresponding feature data. Through a large amount of historical data, the SVM model can learn the unique features of different fault types, such as pipe blockage, pump failure, valve leakage, and heat exchanger efficiency degradation. The training process improves the model's ability to distinguish fault types by optimizing the SVM's classification boundary.

[0058] A trained SVM model can output two important pieces of information: fault type and fault probability. Fault types mainly include pipe blockage, pump failure, valve leakage, and heat exchanger efficiency degradation. Each fault type has its specific manifestations and characteristics, and the SVM model can quickly pinpoint the problem.

[0059] Simultaneously, the SVM model also outputs a failure probability, which is the likelihood of each failure type occurring. This probability is calculated using the model's decision function, which is based on the distance between the feature vector and the classification hyperplane. This distance represents the proximity of the sample point to the failure type boundary, thus determining the probability that the sample belongs to a certain failure type.

[0060] The fault diagnosis results are updated in real time. As the system's operating status changes, new data is continuously input into the model, and the SVM model recalculates and diagnoses. Real-time updates help to promptly identify new faults or adjust the judgment of known faults.

[0061] In addition, the fault diagnosis results are compared with a historical fault database. This database contains past fault records and corresponding feature data. By comparing these records, the system can identify whether the current fault is similar to a previous type of fault, further improving diagnostic accuracy. If the current fault is similar to a historical fault, the model can utilize historical data to improve the recognition rate of that type of fault, avoiding misdiagnosis due to unknown characteristics of new faults.

[0062] By combining support vector machine models with multi-dimensional time series data features, the fault diagnosis process can efficiently and accurately identify fault types in the heating system and provide diagnostic results in real time. This not only improves the efficiency of fault diagnosis but also enhances the intelligence level of the heating system, helping to take timely measures to prevent further deterioration of the fault and ensure the stable operation of the system.

[0063] In one possible implementation, the system first constructs a dynamic model that describes the operating state of the heating system based on the principles of mass and energy conservation. Mass conservation ensures that the total mass of hot water in the system remains constant as it flows between the various components, while energy conservation ensures a balance between the input and output of thermal energy. The model's inputs include: Heat load forecast: Based on the load forecast results from the previous steps, this provides the system's heat demand for a future period of time.

[0064] Fault diagnosis results: from the fault diagnosis module, providing the type and probability of the current fault in the system, helping the model to consider the impact of the fault.

[0065] Current operating parameters: including real-time system operating status such as temperature, pressure, and flow rate.

[0066] The model calculates the setpoints for the control variables based on these input data. The control variables include pump speed, valve opening, and heating power, etc.

[0067] The objective of optimal control is to maximize the system's energy efficiency while minimizing potential failure risks. To this end, the optimization solver uses the following objective function: Maximizing energy efficiency: The goal is to maximize the match between actual heating supply and demand load by adjusting control variables. In other words, it is to ensure that the difference between the system's heating supply and demand is minimized, thereby reducing energy waste.

[0068] Fault risk minimization: The goal is to reduce the probability of system failure. The optimization solver, based on fault diagnosis results, weights and sums the risk values ​​of various fault types as a measure of fault risk, and adjusts system operating parameters as much as possible to reduce the probability of failure.

[0069] To find the optimal settings for the control variables, the optimization solver uses gradient descent for iterative optimization. Gradient descent finds the points where the objective function is minimized (or maximized) by calculating the partial derivatives of the objective function with respect to the control variables. The control variable settings are updated iteratively multiple times until the objective function converges to the optimal solution.

[0070] Once the optimal values ​​of the control variables are calculated, the system sends them to the actuators. The actuators then adjust various parameters of the heating system based on these control variables, including: Pump speed: The flow rate is controlled by adjusting the pump speed, thereby regulating the flow rate of hot water.

[0071] Valve opening: Adjusting the valve opening controls the flow rate of hot water.

[0072] Heating power: Adjust the power output of the heating equipment to ensure that the heat load requirements are met.

[0073] The actuator controls various aspects of the heating system by adjusting these parameters, making it operate as efficiently as possible.

[0074] The adjusted control results are fed back to the data processing unit, achieving closed-loop control. This means that the system adjusts its operating status in real time based on the feedback information to ensure that the heating system is always in optimal operating condition.

[0075] By combining model predictive control and an optimization solver, this method effectively improves the energy efficiency and safety of heating systems, reduces failure risks, and enhances system stability and adaptability. This optimization control method is not only crucial for the daily management of heating systems but also significantly extends equipment lifespan and reduces maintenance costs.

[0076] In one possible implementation, an adaptive learning rate adjustment strategy is employed during LSTM model training to improve training performance and accelerate convergence. Specifically, the learning rate is dynamically adjusted based on changes in the training error. When the training error decreases, the learning rate is gradually reduced to avoid excessively large step sizes that could cause model oscillations or instability; conversely, when the error increases, the learning rate is appropriately increased to accelerate convergence and prevent the model from getting trapped in local optima. This dynamic adjustment of the adaptive learning rate is achieved through an exponential decay method. This method decreases exponentially in each iteration based on the training error, allowing for more precise parameter adjustments during training, especially in later stages, resulting in smoother and more stable convergence.

[0077] Because heating systems exhibit a certain degree of inertia—meaning their state changes relatively gradually, and the impact of external environmental factors (such as temperature changes) on the system has a time lag—the training data for the LSTM model is generated using a sliding window method. The size of the sliding window is determined based on the system's inertia and changes in the external environment, aiming to capture data over a sufficiently long period to identify the system's long-term dependencies and trends. Larger windows are better suited for extracting information about long-term dependencies, while smaller windows are more suitable for capturing short-term dynamic changes. Therefore, the window setting must consider not only the length of historical data but also the characteristics of environmental fluctuations to ensure the model can fully understand and learn the operating patterns of the heating system.

[0078] To evaluate the generalization ability of LSTM models, hold-out validation is typically used. Hold-out validation divides the entire dataset into two subsets: a training set for training and a test set for testing. Evaluating the model on the test set effectively verifies whether the model can accurately predict on unseen data, avoiding performance degradation due to overfitting the training data. This method helps ensure that the trained model not only performs well on the training set but also copes with constantly changing heating demands and environmental conditions in reality.

[0079] Overfitting can occur during the training of Long Short-Term Memory (LSTM) networks, especially when the amount of data is small or the model complexity is high. To prevent this, LSTM models introduce a regularization strategy. Specifically, a regularization term is added to the model's loss function to help reduce the risk of overfitting. The regularization term is designed based on the L2 norm of the model weights, meaning that regularization, by penalizing the sum of squared weights, forces the model to maintain as small a weight value as possible during training, thereby reducing model complexity and overfitting to the training data.

[0080] This invention enables the LSTM model to possess powerful learning capabilities and excellent generalization ability in intelligent load forecasting, accurately predicting load changes in heating systems and providing a reliable basis for subsequent fault diagnosis and optimized control. This not only improves the operating efficiency of the heating system but also enhances its stability and adaptability.

[0081] In one possible implementation, Support Vector Machines (SVMs) are a commonly used supervised learning model for classification, capable of effectively handling both linear and nonlinear problems. In heating systems, the features of the data are typically highly nonlinear; therefore, kernel functions are used to map the data from a low-dimensional space to a high-dimensional space, thus addressing the nonlinearity. Specifically, Radial Basis Functions (RBFs) are used, capable of capturing complex patterns and performing well in many practical problems. The parameters of the RBFs (such as their width) determine the shape of the mapped feature space, and their optimization is crucial. To optimize these parameters, a grid search method is employed, systematically searching through multiple possible parameter combinations to find the parameters best suited for the dataset, thereby improving the model's classification accuracy.

[0082] Support Vector Machines (SVMs) separate data from different classes by determining an optimal classification hyperplane. During training, the model needs to solve a convex optimization problem. The goal of this problem is to find a hyperplane that maximizes the margin between the two classes of data points, thereby improving classification accuracy. To solve this optimization problem, the Sequential Minimum Optimization (SMO) algorithm is used. This algorithm approximates the optimal solution by progressively optimizing the Lagrange multipliers of a pair of support vectors. It can effectively handle large-scale datasets and avoids the excessive computational complexity that can occur with traditional optimization methods.

[0083] Real-time performance is a crucial indicator in fault diagnosis of heating systems. To improve diagnostic efficiency, parallel computing techniques are employed, particularly in the feature extraction stage. Feature extraction involves transforming and processing raw data to extract information meaningful for fault diagnosis. Since this process may involve substantial data processing and computation, multi-threading techniques can significantly improve processing speed. Specifically, multi-threading, based on task decomposition and data partitioning strategies, breaks down large tasks into multiple subtasks and assigns them to different threads for parallel processing. This improves computational efficiency, reduces response time, and ensures high real-time performance of the fault diagnosis system in practical applications.

[0084] In intelligent fault diagnosis, besides determining whether a fault exists, it is also necessary to assess its severity. The output of a Support Vector Machine (SVM) includes not only the fault category information but also its severity. To calculate the severity, the system uses two metrics: the magnitude of the feature vector and the classification confidence score. The magnitude of the feature vector reflects the distance of the data point from the classification hyperplane; the greater the distance, the more severe the fault. The classification confidence score measures the model's confidence in the classification result; the higher the confidence score, the greater the probability of the fault, and usually corresponds to a more severe fault.

[0085] By combining the nonlinear feature processing, optimized calculation process, real-time computing capabilities, and fault severity assessment of support vector machines, the entire system can more accurately and efficiently complete intelligent load prediction and fault diagnosis tasks, thereby improving the stability and safety of the heating system.

[0086] In one possible implementation, the dynamic model of the model predictive control method describes the dynamic behavior of the heating system through state-space equations. The state-space equations comprise two important components: the mass balance equation and the energy balance equation. The mass balance equation describes the mass flow of various fluids in the system, ensuring the conservation of mass within the system; while the energy balance equation tracks the energy conversion and transfer within the system, including heat input, output, and storage. Through these two equations, the dynamic model can reflect the state changes of the system at different points in time, providing accurate state predictions for optimizing the control process.

[0087] Extending equipment lifespan is a crucial objective in optimized control processes. Long-term reliable operation of equipment is essential for the stability of the heating system. To achieve this, the system extends equipment lifespan by constraining the smoothness of control variables. Specifically, the rate of change of control variables (such as valve opening and pump speed) must be limited to prevent frequent and drastic adjustments to the equipment. By constraining the smoothness of control inputs, equipment load fluctuations are reduced, effectively slowing down equipment wear and aging, and extending its service life.

[0088] The core task of an optimization solver is to solve optimization problems with constraints. In this process, the interior-point method is an effective mathematical optimization technique widely used to handle complex constrained problems. Through the interior-point method, the system can transform a constrained optimization problem into an unconstrained one, thus simplifying the solution process. Specifically, the interior-point method introduces a barrier function to "penalize" solutions that do not meet the constraints, ensuring that the solver always avoids these infeasible regions when searching for the optimal solution, thereby guaranteeing the stability and convergence of the solution process.

[0089] To cope with external disturbances and rapidly changing operating conditions in heating systems, the frequency of control command generation needs to be dynamically adjusted. Based on real-time data acquisition and prediction error feedback, the system flexibly adjusts the frequency of control command generation according to the system's response time and the degree of external disturbance. For example, when the system is stable, the frequency of control command generation can be lower to save computing resources; while when the system experiences significant disturbances or faults, the frequency can be increased to ensure rapid system response and timely adjustments. This dynamic adjustment process, by acquiring real-time data on system status and changes in the external environment and combining this with model prediction errors, ensures the rapid adaptability and high efficiency of the control system.

[0090] Through the embodiments of the present invention, the intelligent load prediction and optimal control of the heating system can more effectively achieve load regulation, fault diagnosis and equipment management, thereby improving the operation efficiency, stability of the heating system and the service life of the equipment, and further reducing the operation and maintenance costs and improving the energy utilization rate.

[0091] In a possible implementation, the load prediction step first predicts the future load demand of the heating system based on historical data and real-time environmental data. The prediction results are input to the fault diagnosis step in real time. At this time, the fault diagnosis module analyzes the current operating state of the system and identifies potential faults or anomalies according to the load prediction results. The results of the fault diagnosis (such as the type and location of the fault) are then transmitted to the optimization control step in real time to provide a basis for subsequent adjustment of the control strategy.

[0092] After the fault diagnosis results are input, the optimization control step adjusts the control strategy based on the fault probability. If the diagnosed fault probability exceeds a predetermined threshold, the system triggers a fault mitigation strategy. These strategies include: Reduce pump speed: By reducing the speed of the pump, the load of the pump is reduced, thereby reducing the system pressure and avoiding overload or damage of the pump.

[0093] Adjust valve opening: According to the fault diagnosis results, automatically adjust the valve opening to optimize the water flow and heat distribution, and avoid local overheating or overcooling.

[0094] Reduce heating power: In case of a fault, reducing the heating power of the heating equipment helps to reduce the system burden and relieve the equipment pressure.

[0095] The entire linkage process is implemented by a data processing unit, which is responsible for real-time data stream processing. The data stream processing adopts a pipeline architecture to ensure seamless connection of the load prediction results, fault diagnosis information and optimization control instructions. The pipeline architecture divides data processing into multiple stages, and each stage processes specific tasks, ensuring that the data processing of each step is independent and parallel, thereby reducing system latency and improving the response speed. The pipeline structure can ensure the rapid flow of data from one module to another, adjust the system control in real time, and ensure the efficient operation of the system.

[0096] System performance monitoring is achieved through a logging and alarm mechanism. The logging can record the execution of each step in detail, including: Prediction accuracy: Record the error and accuracy of the load prediction, and analyze the reliability of the prediction results.

[0097] Diagnosis accuracy: Record the accuracy of the fault diagnosis to ensure the consistency between the diagnosis results and the actual fault conditions.

[0098] Control efficiency indicators: These record the effects of implementing control strategies, such as energy efficiency optimization and system stability.

[0099] In addition, when the system malfunctions, the alarm mechanism will be triggered immediately to remind operators to pay attention to potential problems and ensure that the system is dealt with in a timely manner.

[0100] Through real-time data transmission and linkage between steps, the system can quickly respond to faults or load changes, implementing effective fault mitigation and optimized control strategies. Fault mitigation strategies can take measures when potential equipment failures occur, preventing equipment damage or excessive wear and tear, effectively extending equipment lifespan. Pipeline architecture and real-time data stream processing improve the system's response speed and processing capacity, reduce control latency, and ensure more stable and efficient system operation. Through logging and alarm mechanisms, the system can continuously monitor the performance of each component, ensuring the system is always in optimal operating condition.

[0101] Example This embodiment is implemented in an actual operating heating system, which includes a heat source plant, a distribution network, and user terminals. The heat source plant uses gas-fired boilers, the distribution network includes pumps, valves, and heat exchangers, and user terminals receive heat through radiators. The system covers an area of ​​approximately 500,000 square meters and serves more than 1,000 households. The implementation period is one heating season (from November of the current year to March of the following year), with data acquisition frequency once per minute and optimized control commands updated every 5 minutes.

[0102] Step 1: Multi-dimensional data collection; This step involves collecting real-time operating and environmental parameters of the heating system. The specific process is as follows: Temperature data is collected by a PT100 temperature sensor, which is installed at the inlet and outlet of the heating network to measure the inlet and outlet water temperatures. The measurement range is 0-150℃, with an accuracy of ±0.1℃.

[0103] Flow data is collected by an electromagnetic flow meter installed on the main pipeline to measure flow rate, with a measurement range of 0-1000m. 3 / h, accuracy ±0.5%.

[0104] Pressure data is acquired through a piezoresistive pressure sensor, which is installed at the pump outlet and before and after the valve to measure pressure data. The measurement range is 0-2.5MPa, with an accuracy of ±0.1%.

[0105] Vibration data is acquired by an accelerometer installed on the housing of the pump and valve. The measurement frequency range is 0-10kHz, and the sensitivity is 100mV / g.

[0106] Energy consumption data is collected through a heat meter installed at the user entrance to measure heat energy consumption data with a measurement accuracy of ±1%.

[0107] Environmental data is collected by temperature and humidity sensors installed in an outdoor shaded area to measure outdoor temperature and humidity. The temperature range is -40℃ to 60℃, and the humidity range is 0-100%RH, with an accuracy of ±0.5℃ and ±3%RH.

[0108] All sensor data is acquired via a data acquisition device (using an NICDAQ-9188 Ethernet chassis, equipped with an analog input module) at a preset sampling frequency. The preset sampling frequency is initially set to 1 time / minute, but is dynamically adjusted according to the operating cycle of the heating system and changes in the external environment. The dynamic adjustment process is based on historical data analysis and real-time load demand: the data processing unit (industrial computer, configured with an Intel i7 processor) calculates the coefficient of variation (the ratio of standard deviation to mean) of the data over the past 24 hours. When the coefficient of variation exceeds 0.1, the sampling frequency is increased to 1 time / 30 seconds; when the coefficient of variation is below 0.05, the sampling frequency is decreased to 1 time / 2 minutes. During data acquisition, a UTC timestamp is appended to each parameter to ensure data synchronization. This step outputs a multi-dimensional time series dataset, stored in CSV format, containing timestamps and all parameter values.

[0109] Step 2: Data preprocessing; This step preprocesses the multi-dimensional time series dataset output from step 1. The specific process is as follows: Data cleaning: A sliding window method is used to identify and remove outliers. The sliding window size is set to 30 data points (corresponding to 30 minutes of data). For each parameter, the mean and standard deviation of the data within the window are calculated, and data points exceeding three times the standard deviation are marked as outliers and deleted. For example, for temperature data, if a data point differs from the window mean by more than three times the standard deviation, that point is deleted.

[0110] Data imputation: Missing data is filled using linear interpolation. Linear interpolation is based on the trend of adjacent data points: for a missing point, the arithmetic mean of the two valid data points before and after it is used as the imputation value. If the missing point is located at the beginning or end of the sequence, it is filled by copying the nearest valid data point.

[0111] Data standardization: A min-max standardization method is used to scale each parameter to between zero and one. The specific formula is: Standardized value = (Original value - Minimum value) / (Maximum value - Minimum value), where the minimum and maximum values ​​are taken from the historical range of the entire dataset (e.g., the historical range of temperature data is 0-150℃). After standardization, all parameters are within the same numerical range, which facilitates model processing.

[0112] Data alignment: A timestamp matching algorithm ensures consistency in the time series of all parameters. The algorithm uses the nearest neighbor interpolation method: based on the timestamp of the traffic data, the timestamps of other parameters are aligned to the nearest timestamp of the traffic data, with a deviation of no more than 1 second. The preprocessed multi-dimensional time series dataset is stored in an SQL database for use in subsequent steps.

[0113] Step 3: Load forecasting; This step, based on the preprocessed multi-dimensional time series dataset output from step 2, uses a long short-term memory network model to predict the heat load forecast sequence for future periods. The specific process is as follows: The Long Short-Term Memory (LSTM) network model structure consists of an input layer, hidden layers, and an output layer. The number of nodes in the input layer depends on the number of input features, including historical temperature data (inlet and outlet water temperatures), historical flow data, historical pressure data, historical energy consumption data, historical environmental data (outdoor temperature and humidity), and time-related data (date type and holiday markers). In the time-related data, date types are categorized as weekdays and weekends, and holiday markers are binary values ​​(1 for holidays, 0 for non-holidays). The hidden layer uses LTM network units, with the number of units dynamically determined: through cross-validation, the model performance is tested on the training set with 50 to 200 units, and the unit number with the smallest root mean square error is selected (in this embodiment, the final number of units is 100). The output layer is a fully connected layer with one node, outputting the predicted heat load value (unit: GJ / h) for the next hour.

[0114] Model Training: The training process uses a historical multi-dimensional time series dataset (data from the past 3 years, totaling approximately 1.5 million records). The dataset is divided into training, validation, and test sets in a 70:15:15 ratio. Backpropagation is used to optimize model parameters. Backpropagation is based on gradient descent to minimize prediction error (the loss function is the root mean square error). The learning rate for gradient descent uses an adaptive learning rate adjustment strategy: the initial learning rate is 0.001, dynamically adjusted every 10 epochs based on changes in the validation set error; if the validation set error does not decrease for 3 consecutive epochs, the learning rate is multiplied by 0.5. The training epochs are set to 100, and the batch size is 32. Overfitting is prevented by adding an L2 regularization term to the loss function, with a regularization coefficient of 0.01.

[0115] Model output: The model input consists of historical data from the past 24 hours (1440 data points), and the output is a sequence of predicted heat load values ​​for the next hour (12 data points, one value every 5 minutes). The predicted values ​​are expressed as heat energy demand per unit time and are updated in real time.

[0116] Step 4: Fault diagnosis; This step, based on the preprocessed multi-dimensional time series dataset output from step 2 and the heat load prediction sequence output from step 3, performs fault diagnosis. The specific process is as follows: Feature extraction: Extracting features from multi-dimensional time series datasets, including: Temperature change rate: Calculate the slope of the temperature difference between the inlet water temperature and the return water temperature using a linear regression method over a 10-minute window.

[0117] Traffic fluctuation range: Calculate the standard deviation of traffic data over a 10-minute window.

[0118] Pressure gradient: Calculate the first-order difference mean of the pressure data over a 10-minute window.

[0119] Vibration spectrum characteristics: Perform fast Fourier transform on the vibration data to extract the main frequency amplitude over a 10-minute window.

[0120] Deviation between predicted and actual heat load: Calculate the absolute difference between the predicted value and the actual energy consumption value output in step 3 over a 10-minute window.

[0121] Support Vector Machine (SVM) Model Training: The model learns patterns of normal and various fault states through training. Training data includes historical fault records (500 fault events from the past two years) and corresponding feature data. Fault types include pipe blockage, pump failure, valve leakage, and heat exchanger efficiency degradation. The SVM model uses radial basis functions (RBFs) as kernel functions. The parameters (γ and C) of the RDFs are optimized using a grid search method: γ takes values ​​of [0.001, 0.01, 0.1, 1], and C takes values ​​of [0.1, 1, 10, 100]. The parameter combination with the highest accuracy is selected on the validation set (γ = 0.1, C = 10 in this embodiment). The classification hyperplane is determined by solving a convex optimization problem using a sequential minimum optimization algorithm. The model outputs the fault type and fault probability. The fault probability is calculated using the SVM model's decision function: the decision function is based on the distance between feature vectors and the classification hyperplane, and the probability value is converted to a value between 0 and 1 using the Plattscaling method.

[0122] Real-time fault diagnosis: Achieved through parallel computation of feature extraction using multi-threading technology (Python multiprocessing library). Data is partitioned for processing, with each thread handling feature extraction for one parameter. Diagnostic results are updated every 5 minutes and compared with a historical fault database to improve diagnostic accuracy. Fault severity is calculated using the modulus of the feature vector and classification confidence: a larger modulus indicates higher severity; when the confidence is below 0.7, the diagnostic result is marked as "pending confirmation".

[0123] Step 5: Optimize control; This step generates an optimized control strategy based on the heat load prediction sequence output in step 3 and the fault diagnosis results output in step 4. The specific process is as follows: Model predictive control method: A dynamic model of the heating system is constructed, which is represented by state-space equations, including mass balance equations and energy balance equations. Mass balance equation: dm / dt=ρ*(Q_in-Q_out), where m is the system mass, ρ is the density, and Q_in and Q_out are the inflow and outflow rates. Energy balance equation: dT / dt=(P_heat-U*A*(T-T_env)) / (m*C_p), where T is the temperature, P_heat is the heating power, U is the heat loss coefficient, A is the surface area, T_env is the ambient temperature, and C_p is the specific heat capacity. Model inputs include predicted heat load, fault diagnosis results (fault type and probability), and current operating parameters (temperature, flow rate, pressure). Model outputs are control variable setpoints, including pump speed (unit: rpm), valve opening (unit: %), and heating power (unit: kW).

[0124] The optimization solver solves for the setpoints of the control variables with the objective functions of maximizing energy efficiency and minimizing fault risk. The objective function is expressed as: J = α*(1-|Q_actual-Q_pred| / Q_pred) + β*(1-Σw_i*P_fault_i), where Q_actual is the actual heating supply, Q_pred is the predicted heat load, P_fault_i is the probability of the i-th fault, w_i is a weighting coefficient (set according to the severity of the fault; in this embodiment, w=0.4 for pipe blockage, w=0.3 for pump failure, w=0.2 for valve leakage, and w=0.1 for heat exchanger efficiency reduction), and α and β are balance coefficients (α=0.7, β=0.3). Maximizing energy efficiency represents the matching degree between actual heating supply and demand heat load, and minimizing fault risk represents the weighted sum of fault probabilities. The optimization solution process uses the gradient descent method to iteratively update the setpoints of the control variables, with an upper limit of 100 iterations, and the convergence condition is that the change in the objective function is less than 0.001. The constraints include the following control variable ranges: pump speed 0-3000 rpm, valve opening 0-100%, and heating power 0-5000 kW. The optimization solution uses the interior-point method to handle the constraints, which transforms the constrained problem into an unconstrained problem through obstacle functions.

[0125] Control command generation: Control variable setpoints are sent to actuators (variable frequency pumps, electric valves, and boiler controllers) to adjust pump speed, valve opening, and heating power. The actuators feed back the control results to the data processing unit, achieving closed-loop control. The frequency of control command generation is dynamically adjusted based on system response time and external disturbances: based on real-time data acquisition and prediction error feedback, the frequency increases to once every minute when the prediction error exceeds 10%; otherwise, it remains once every 5 minutes.

[0126] The coordinated process of steps 3, 4, and 5 In this embodiment, steps 3, 4, and 5 are linked in real time: The load prediction results are input into the fault diagnosis step every 5 minutes as part of feature extraction.

[0127] The fault diagnosis results are input into the optimization control step every 5 minutes to adjust the fault risk weights in the objective function.

[0128] The optimized control steps adjust the control strategy based on the failure probability: when the failure probability exceeds a threshold of 0.8, the optimized control steps prioritize the execution of failure mitigation strategies. Failure mitigation strategies include reducing the pump speed to a safe range (e.g., reducing it by 20%), adjusting valve openings to balance flow (e.g., increasing the opening of leaking valves by 10%), and reducing heating power to the minimum required level (e.g., reducing it by 15%).

[0129] The entire linkage process achieves real-time data stream processing through the data processing unit, using a pipeline architecture: each step is an independent process, and data is transmitted through a message queue (Redis) to ensure seamless data transmission between steps.

[0130] System performance monitoring is achieved through logging and alarm mechanisms: logging includes indicators such as prediction accuracy (root mean square error), diagnostic accuracy (confusion matrix), and control efficiency (energy saving rate); when any indicator is abnormal (such as prediction error exceeding 15%), an SMS alarm is triggered.

[0131] Through the implementation of this embodiment, the load prediction error of the heating system is reduced from 15% to 5% by traditional methods, the fault diagnosis delay is reduced from hours to minutes, and the control energy efficiency is improved by 20%, fully demonstrating the effectiveness and ingenuity of the present invention.

[0132] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0133] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent load prediction, fault diagnosis, and optimized control of heating systems, characterized in that, Includes the following steps: Step 1: Multi-dimensional data acquisition. Collect real-time operating parameters and environmental parameters of the heating system, including temperature data, flow data, pressure data, vibration data, energy consumption data, and environmental data. All data are collected by sensors and data acquisition devices at a preset sampling frequency and transmitted to the data processing unit to generate a multi-dimensional time series dataset. Step 2: Data preprocessing. The multi-dimensional time series dataset output in Step 1 is subjected to data cleaning, data imputation, data standardization and data alignment. Data cleaning identifies and removes outliers, data imputation uses interpolation methods to fill in missing data, data standardization scales the data to the same numerical range, and data alignment ensures that the timestamps of all parameters are consistent, generating a preprocessed multi-dimensional time series dataset. Step 3: Load forecasting. Based on the preprocessed multi-dimensional time series dataset output from Step 2, a long short-term memory network model is used to predict the heat load forecast sequence for future periods. The inputs of the long short-term memory network model include historical operating parameters, historical environmental parameters, and time feature data. Step 4: Fault diagnosis. Based on the preprocessed multi-dimensional time series dataset output in Step 2 and the heat load prediction value sequence output in Step 3, extract the features from the multi-dimensional time series dataset and input the features into the support vector machine model. The support vector machine model outputs the fault type and fault probability. Step 5: Optimize control. Based on the heat load prediction sequence output in Step 3 and the fault diagnosis results output in Step 4, use the model predictive control method to generate control commands. The model predictive control method constructs a dynamic model of the heating system and optimizes the setpoints of control variables, including pump speed, valve opening, and heating power. The optimization objectives include maximizing energy efficiency and minimizing fault risk.

2. The intelligent load prediction, fault diagnosis, and optimized control method for a heating system as described in claim 1, characterized in that, The specific process of multi-dimensional data collection in step 1 includes: Temperature data is collected from the inlet and return water temperatures of the heating network via temperature sensors. Flow data is collected from the heating network through flow sensors; Pressure data is collected from the heating network via pressure sensors. Vibration data is collected from pumps and valves using vibration sensors; Energy consumption data is collected through energy metering devices to collect thermal energy consumption data; Environmental data is collected through environmental sensors, which gather outdoor temperature and humidity data. All sensor data are collected by the data acquisition device at a preset sampling frequency and transmitted to the data processing unit. During the data acquisition process, each parameter is timestamped to ensure data synchronization. The preset sampling frequency is dynamically adjusted according to the operating cycle of the heating system and changes in the external environment. The dynamic adjustment process is based on historical data analysis and real-time load demand, and the optimal sampling interval is calculated by the data processing unit.

3. The intelligent load prediction, fault diagnosis, and optimized control method for a heating system as described in claim 1, characterized in that, The specific process of data preprocessing in step 2 includes: Data cleaning uses a sliding window method to calculate the mean and standard deviation of each parameter, and data points exceeding three times the standard deviation are marked as outliers and deleted. Data imputation uses linear interpolation to fill in missing data, which is based on the trend of adjacent data points. Data standardization employs a min-max standardization method, transforming each parameter to a numerical range between zero and one. Data alignment ensures the consistency of the time series of all parameters through a timestamp matching algorithm, which uses the nearest neighbor interpolation method to align data points with different sampling frequencies. The preprocessed multidimensional time series dataset is stored in a database for use in subsequent steps.

4. The intelligent load prediction, fault diagnosis, and optimized control method for a heating system as described in claim 1, characterized in that, The specific process of load forecasting in step 3 includes: The inputs to the Long Short-Term Memory Network model include historical temperature data, historical flow data, historical pressure data, historical energy consumption data, historical environmental data, and time feature data, including date type and holiday markers. Long Short-Term Memory Network Model learns non-linear relationships in data through the training process. The training process uses historical multi-dimensional time series datasets and optimizes model parameters through the backpropagation algorithm, which minimizes prediction error based on the gradient descent method. The model output is a sequence of predicted heat load values ​​for a specified future time period, expressed as heat energy demand per unit time. The number of hidden layer nodes in the Long Short-Term Memory (LSTM) network model is dynamically determined based on the data dimensionality and prediction accuracy requirements. The optimal number of nodes is selected through cross-validation.

5. The intelligent load prediction, fault diagnosis, and optimized control method for a heating system as described in claim 1, characterized in that, The specific process of fault diagnosis in step 4 includes: Features were extracted from the multi-dimensional time series dataset, including temperature change rate, flow fluctuation amplitude, pressure gradient, vibration spectrum characteristics, and the deviation between predicted and actual heat load values. The support vector machine model learns patterns of normal states and various fault states through the training process. The training data includes historical fault records and corresponding feature data. The support vector machine model outputs fault types and fault probabilities. Fault types include pipe blockage, pump failure, valve leakage, and heat exchanger efficiency degradation. The failure probability is calculated using the decision function of the support vector machine model, which is based on the distance between feature vectors and the classification hyperplane. The fault diagnosis results are updated in real time and compared with the historical fault database to improve the accuracy of diagnosis.

6. The intelligent load prediction, fault diagnosis, and optimized control method for a heating system as described in claim 1, characterized in that, The specific process of optimizing control in step 5 includes: The model predictive control method constructs a dynamic model of the heating system. The dynamic model is based on the principles of mass conservation and energy conservation. The inputs include the predicted heat load, fault diagnosis results, and current operating parameters, and the output is the setpoint of the control variables. The optimization solver solves for the setpoints of the control variables with the objective functions of maximizing energy efficiency and minimizing failure risk. In the objective function, maximizing energy efficiency is represented by the matching degree between actual heating supply and demand heat load, and minimizing failure risk is represented by the weighted sum of failure probabilities. The optimization process uses gradient descent to iteratively update the control variable settings until the objective function converges. The control variable setpoints are sent to the actuator to adjust the pump speed, valve opening, and heating power; the actuator feeds back the control results to the data processing unit to achieve closed-loop control.

7. The intelligent load prediction, fault diagnosis, and optimized control method for a heating system as described in claim 4, characterized in that, The training process for a Long Short-Term Memory (LSTM) network model also includes: An adaptive learning rate adjustment strategy is used, in which the adaptive learning rate is dynamically adjusted according to changes in training error, and the dynamic adjustment process is based on the exponential decay method. The input data for the Long Short-Term Memory Network model is used to generate training samples through a sliding window method. The size of the sliding window is determined based on the inertial characteristics of the heating system and changes in the external environment. Model validation uses the hold-out method, which divides the dataset into a training set and a test set, with the test set used to evaluate the model's generalization ability. Overfitting is prevented by adding a regularization term to the loss function, which is based on the L2 norm of the weights.

8. The intelligent load prediction, fault diagnosis, and optimized control method for a heating system as described in claim 5, characterized in that, The training process of a support vector machine model also includes: Kernel functions are used to handle nonlinear characteristics. Radial basis functions are selected as kernel functions, and the parameters of the radial basis functions are optimized through a grid search method. The classification hyperplane of the support vector machine model is determined by solving a convex optimization problem, which is solved using the sequential minimum optimization algorithm. The real-time nature of fault diagnosis is achieved through parallel computing of feature extraction. Parallel computing uses multi-threaded processing technology, which is based on task decomposition and data partitioning. The output of the support vector machine model also includes the severity of the fault, which is calculated using the magnitude of the feature vector and the classification confidence.

9. The intelligent load prediction, fault diagnosis, and optimized control method for a heating system as described in claim 6, characterized in that, The dynamic models of model predictive control methods also include: The dynamic behavior of a heating system is represented by state-space equations, which include mass balance equations and energy balance equations. The objective function of the optimizer also includes extending equipment lifespan, which is achieved through smoothness constraints on control variables. Smoothness constraints are expressed as limits on the rate of change of control variables. The optimization solution process uses the interior point method to handle constraints, which transforms the constrained problem into an unconstrained problem through obstacle functions. The frequency of control command generation is dynamically adjusted based on system response time and external disturbances. The dynamic adjustment process is based on real-time data acquisition and prediction error feedback.

10. The intelligent load prediction, fault diagnosis, and optimized control method for a heating system as described in claim 1, characterized in that, The linkage process of steps 3, 4, and 5 includes: Load forecasting results are input to the fault diagnosis step in real time, and fault diagnosis results are input to the optimization control step in real time. The optimized control steps adjust the control strategy based on the failure probability. When the failure probability exceeds the threshold, the optimized control steps prioritize the execution of the failure mitigation strategy. Fault mitigation strategies include reducing pump speed, adjusting valve opening, and reducing heating power; The entire linkage process achieves real-time data stream processing through the data processing unit. The data stream processing uses a pipeline architecture, which ensures seamless data transmission between steps. System performance monitoring is achieved through logging and alarm mechanisms. The logs include indicators such as prediction accuracy, diagnostic accuracy, and control efficiency.

Images