HVAC digital twin energy efficiency prediction method based on lstm and polynomial fitting

Abstract

This invention discloses an HVAC digital twin energy efficiency prediction method based on LSTM and polynomial fitting. The method first collects and preprocesses multi-dimensional operating data of the HVAC system, then constructs a component-level high-order polynomial fitting model containing aging and pressure loss coupling terms. This model is combined with a bidirectional LSTM model with an attention mechanism to predict energy consumption and COP for the next 1-4 hours. A digital twin is built based on the component model to achieve dual-modal simulation. The NSGA-II algorithm is used to complete multi-objective optimization and adaptively adjust weights to generate optimal control commands. Simultaneously, a deviation-triggered model incremental update and rollback mechanism is established. This invention achieves high-fidelity simulation and accurate prediction, forming a closed-loop optimization system, significantly improving system COP, reducing energy consumption and grid peak-valley differences, and is adaptable to all operating conditions and complex operating scenarios.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology of heating, ventilation and air conditioning (HVAC), specifically involving an HVAC digital twin energy efficiency prediction method based on LSTM and polynomial fitting. It is applicable to central air conditioning systems that include components such as chiller units, chilled water pumps, cooling water pumps, and cooling towers. It can cover all operating scenarios from low load to ultra-high load and is suitable for complex scenarios such as variable frequency / fixed frequency mixed water pump configuration, equipment aging, and transitional operating conditions. Background Technology

[0002] As a core component of building energy consumption, the energy efficiency of HVAC systems directly impacts building energy consumption and operating costs. Currently, HVAC systems face numerous technical bottlenecks in modeling, prediction, and optimization, making it difficult to meet the demands for higher efficiency and intelligence. (1) Insufficient accuracy of component models: The performance curves of core equipment such as chillers, pumps, and cooling towers exhibit strong nonlinear characteristics and are significantly affected by factors such as equipment aging and pipeline pressure loss. Existing technologies mostly use linear, quadratic, or cubic low-order models for fitting, without considering the above-mentioned coupling effects, and cannot accurately characterize the operating characteristics of the equipment in the full range of operating conditions. The model determination coefficient ( The low energy efficiency results in large discrepancies between energy consumption calculations and simulations, making it difficult to support precise energy efficiency optimization.

[0003] (2) Lack of time-series forecasting capability: Traditional HVAC system models can only calculate the current energy consumption status based on real-time data, lacking the ability to predict future energy consumption trends. Since building cooling load, outdoor environmental parameters, etc. are all dynamically changing, lagging control can easily lead to a mismatch between system operation and actual demand, resulting in energy waste or poor cooling effect. At the same time, existing models do not fully explore the long-term dependencies in historical data, and the prediction accuracy is difficult to meet engineering requirements.

[0004] (3) Insufficient synergy between simulation and optimization: In existing technologies, system simulation and energy efficiency optimization are mostly independent links, and a closed-loop system of "prediction-simulation-optimization" has not been formed. Even if some solutions involve optimization, they are mostly based on a single objective or static operating conditions, without considering key indicators such as the difference between peak and valley loads of the power grid and the energy consumption fluctuations under transitional operating conditions. They are difficult to adapt to the multi-objective optimization needs under complex dynamic loads and cannot achieve optimal energy efficiency throughout the entire life cycle of the system.

[0005] (4) Insufficient model adaptability: The existing model does not consider the parameter drift problem under scenarios such as equipment maintenance and changes in load interval ratio, and lacks an adaptive update mechanism. After long-term operation, the model accuracy will decrease significantly and cannot continuously provide reliable support for optimized control.

[0006] In summary, existing HVAC system technologies lack high-precision model support, high-accuracy time-series prediction capabilities, complete closed-loop optimization mechanisms, and adaptive update capabilities. There is an urgent need for an innovative approach that integrates physical mechanisms and data-driven methods to solve the problems of high-fidelity simulation, forward-looking energy efficiency prediction, and autonomous optimization control under all operating conditions. Summary of the Invention

[0007] (a) Purpose of the invention This invention aims to overcome the shortcomings of existing technologies and provide an HVAC digital twin energy efficiency prediction method based on LSTM and polynomial fitting. This invention employs a dual-model collaborative coupling structure of a component-level polynomial fitting surrogate model and an LSTM time-series prediction model to construct a hybrid digital twin system integrating high-precision physical mechanisms and deep learning algorithms. Under the premise of meeting the predicted cooling load demand, it achieves high-fidelity simulation of the HVAC system from instantaneous to year-round scales, accurately predicting future energy consumption and energy efficiency trends. Furthermore, it generates predictive control commands through a multi-objective optimization algorithm, ultimately achieving the goals of minimizing total system energy consumption, maximizing overall energy efficiency, minimizing the peak-valley load difference of the power grid, and minimizing energy consumption fluctuations during transitional operating conditions. Simultaneously, it improves system operational stability and engineering practicality.

[0008] (II) Terminology and Calculation Methods To clarify the objective function and output indicators of this invention, the following terminology and calculation methods are standardized: all indicators are time-series values ​​in the prediction time domain, and hourly indicator data for the next 1-4 hours are output in an hourly dimension: 1. Total system energy consumption: The sum of the power consumption of each device in the HVAC system, specifically including the hourly cumulative value of power consumption of chiller unit, chilled water pump, cooling water pump, cooling tower fan, and spray pump, in kW.

[0009] 2. Overall COP (Coefficient of Performance): The ratio of the total cooling capacity of an HVAC system to the total energy consumption of the system, including the power consumption of auxiliary equipment such as pumps and towers. It is a core indicator reflecting the overall energy efficiency of the system and has no unit. The total cooling capacity is the sum of the hourly cooling capacity of the chiller units and is in kW (or RT, 1RT = 3.517kW).

[0010] 3. Prediction error: The "mean absolute percentage error (MAPE)" is used to calculate the average absolute percentage error between the predicted and actual values ​​of energy consumption and energy efficiency ratio for the next 4 hours. It is the core indicator for measuring the accuracy of the prediction model.

[0011] 4. Simulation error: The mean absolute percentage error (MAPE) is used to calculate the average absolute percentage error between the total system power consumption and integrated COP output by the digital twin simulation and the measured values ​​on site. It is the core indicator for measuring the accuracy of the simulation model.

[0012] (II) Technical Solution The core technical solution of this invention includes five key steps: data acquisition and preprocessing, LSTM-based energy consumption time series prediction, digital twin dynamic simulation and multi-objective optimization, optimal control command generation and issuance, and model adaptive updating. These steps are closely linked to form a complete closed-loop system of "perception-prediction-simulation-decision-execution-update". The technical solution process is as follows: Figure 1 As shown.

[0013] Step 1: Data Acquisition and Preprocessing Historical and real-time operational data of the HVAC system are collected, and outlier removal and standardization preprocessing are performed on the data to obtain the input dataset for modeling and prediction.

[0014] (1) Data collection scope The operational data includes at least equipment operating status data, environmental data, and load-related data. Specifically, the data collection scope covers these three core categories, encompassing all dimensions of system operation, external environment, and target evaluation. The data sampling interval is one hour. Historical operating parameters: Data collected at preset time intervals, including cooling load, chiller start / stop status, pump and fan operating frequency, chilled water supply and return temperatures, cooling water inlet and return temperatures, pump operating mode, and pipeline pressure loss. External environmental data: Outdoor dry-bulb temperature, wet-bulb temperature, solar radiation intensity, and wind speed collected at preset time intervals; Target variables: total system power consumption, overall energy efficiency ratio (COP), chiller power consumption, water pump power consumption, and cooling tower fan power consumption.

[0015] (2) Data preprocessing process Outlier removal: A combination of the 3σ criterion and the isolated forest algorithm is used to accurately identify two types of outlier data caused by equipment failure and sensor malfunction, with a removal accuracy of ≥98%. Missing value repair: Linear interpolation is used to fill in a small number of missing data to ensure the continuity of data sequence. Standardization: The maximum-minimum normalization method is used to map all data to the [0,1] interval to eliminate the influence of units. The standardization formula is: in, This is the original data. and These are the minimum and maximum values ​​of the data sequence, respectively. Time series feature engineering: Extract the trend, periodic and abrupt change features of the load sequence, decompose them through wavelet transform, and then fuse them with the preprocessed data to form the model input feature set.

[0016] Step 2: Energy Consumption Time Series Prediction Based on LSTM Based on the preprocessed time-series operating data, a Long Short-Term Memory (LSTM) neural network prediction model is constructed. Using multi-dimensional operating data within a preset historical time window as input, the model performs rolling predictions on the total energy consumption and comprehensive energy efficiency ratio (COP) of the HVAC system within a preset prediction time domain, and outputs a time-series prediction sequence of energy consumption and energy efficiency indicators. The preset historical time window is the previous 24 hours, and the preset prediction time domain is the next 1-4 hours.

[0017] The hidden layers of the Long Short-Term Memory (LSTM) neural network prediction model employ a multi-layer bidirectional LSTM structure, with an attention mechanism layer. The output layer is used to output the time-series prediction results of the system's total energy consumption and overall energy efficiency ratio (COP) within a preset future time domain. The LSTM prediction model is trained using the Adam optimizer, and parameters are optimized by combining a dynamic learning rate adjustment strategy and an early stopping mechanism. The specific model design and training parameters are as follows: (1) Model input and output Input layer: 48 neurons, corresponding to the preprocessed data of the first 24 hours and 24-dimensional temporal features; Output layer: 8 neurons, corresponding to the total system energy consumption over the next 1-4 hours ( ) and integrated COP ( Timing value.

[0018] (2) Network structure design The structure adopts a 3-layer bidirectional LSTM hidden layer + attention mechanism layer, where each bidirectional LSTM layer contains 80 neurons, and the attention mechanism layer is used to strengthen the weight of key temporal features and improve prediction accuracy.

[0019] (3) Model training parameters Optimizer: Adam optimizer, initial learning rate 0.001, decaying by 10% every 20 iterations; Number of iterations: 100, batch size = 32; Early stopping mechanism: Training is stopped if the validation set loss increases for 5 consecutive iterations to avoid overfitting; Cross-validation: A cross-validation mechanism is adopted, with 60% of the data used for training, 25% for validation, and 15% for testing, to improve the model's generalization ability.

[0020] (4) Model accuracy index The average forecast error (MAPE) for energy consumption and integrated COP in the next 4 hours is ≤3.2%, with the forecast error for energy consumption in summer ≤3.4% and in winter ≤3.0%, and the forecast error for integrated COP in summer ≤3.2% and in winter ≤2.8%.

[0021] (5) Time-series load-equipment status mapping Based on the LSTM-predicted 24-hour load sequence, a mapping relationship between the load change rate and the optimal state of the equipment is constructed: ,in For the load change rate, It includes parameters such as unit start-up and shutdown status, water pump frequency, number of cooling tower fans and spray intensity, to achieve precise matching between dynamic load and equipment status.

[0022] Step 3: Digital Twin Dynamic Simulation and Multi-Objective Optimization The predicted future cooling load demand from step 2 is used as the driving input and fed into the HVAC digital twin constructed based on the component-level polynomial fitting surrogate model to dynamically simulate the future operating state of the system. A multi-objective optimization algorithm is then used to solve the system operating parameters to obtain an optimized control scheme that meets the preset constraints. The preset constraints are to meet the predicted cooling load demand.

[0023] I. Component-level polynomial fitting surrogate model: For core HVAC system equipment (chillers, pumps, and cooling towers), a high-precision polynomial fitting surrogate model is constructed by combining measured data and physical mechanisms. This model explicitly describes the coupling relationship between the equipment's "flow rate-temperature-speed / frequency-energy consumption / heat exchange". Coupled correction terms such as aging correction and pressure loss are introduced, and the input and output variable sets of each model are clarified. This corrects for long-term performance drift of the equipment and provides a high-fidelity component model foundation for digital twin simulation.

[0024] (1) Chiller unit model Model type: Fourth-order bivariate polynomial fitting model including equipment aging correction term; model determination coefficient. ; Input variables: Partial load factor (PLR), cooling water inlet temperature ( ), cumulative operating time of equipment ( ); Output variables: Chiller COP, Chiller power consumption; Coupling correction item: Equipment aging correction item ,in Equipment aging factor ( ), obtained by fitting the equipment's historical maintenance records with performance degradation data, and its physical meaning is the performance degradation coefficient of the equipment with cumulative running time; Model expression: in, The coefficients are polynomial fitting coefficients, which are calculated using MATLAB tools with a "piecewise fitting + global calibration" strategy (fitting separately according to the PLR ​​intervals of 10%-30%, 31%-70%, and 71%-100%, and then calibrated by global least squares method).

[0025] (2) Water pump model For each chilled water pump and cooling water pump (including variable frequency / fixed frequency pumps), establish a fourth-order univariate polynomial model including pipeline pressure loss coupling terms. The model determination coefficients are... ; Model types: flow-head model, flow-power model, flow-efficiency model, frequency conversion switching loss model; Input variable: Pump operating frequency ( Pipeline pressure loss ); Output variable: Pump head ( ), water pump power consumption ( ), pump efficiency ( ), frequency conversion switching loss ( ); Coupling correction term: Pipeline pressure loss coupling term ( , ),in For pressure loss coefficient, The power loss coefficient is obtained by fitting measured data. Its physical meaning is the additional impact of pipeline pressure loss on the pump head and power. Model expression: 1. Flow-head model: 2. Flow-power model: 3. Flow-efficiency model: 4. Frequency conversion switching loss model: in, , , , The coefficients are obtained by fitting measured data; the fixed-frequency water pump model adopts a simplified fourth-order polynomial form to uniformly describe its operating characteristics.

[0026] (3) Cooling tower model Model type: A thermodynamic model based on the principle of energy conservation and the calculation method of air enthalpy-humidity diagram, combined with wind speed correction mechanism and wind speed-spray intensity coupling relationship, to achieve accurate simulation of cooling tower operation status; Input variables: cooling water flow rate, cooling water inlet / outlet temperature, outdoor wet-bulb temperature ℃), wind speed ( (m / s) spray intensity (L / m) 2 ·h); Output variables: airflow rate, cooling tower fan power consumption, spray pump power consumption, cooling tower outlet water temperature; Coupling correction term: Wind speed correction term Spray intensity coupling term The physical meanings are the influence of wind speed on air volume flow rate and the influence of spray intensity on the total heat release on the water side, respectively. Core calculation process: 1. Enthalpy Calculation: Calculate the enthalpy of the cooling water inlet / outlet and the total heat release on the water side based on the physical property functions; ; .

[0027] 2. Mass flow rate ratio calculation: Based on the law of conservation of energy, the heat released by the water side is equal to the heat absorbed by the air side. Calculate the mass flow rate ratio of water to air. ; 3. Iterative Verification: Using the outdoor wet-bulb temperature as the initial condition, iteratively calculate the air outlet temperature and enthalpy, defining... This ensures that the heat balance error is within the allowable range. 4. Actual flow rate and power consumption calculation: Considering wind speed correction of air volume flow rate, determine the number of operating fans based on the rated air volume of a single fan, and calculate the total power consumption of the cooling tower in combination with the spray intensity.

[0028] in air density, Wind speed; This represents the power consumption coefficient of the spray pump.

[0029] Determining the number of wind turbines: Calculation factors ( (Rated air volume of a single fan), when hour, Adjusted by 1.1 times, according to The range of values ​​determines the number of fans. : II. Digital Twin Simulation Operation Mode: The digital twin simulation operation includes two modes to adapt to different simulation needs: (1) Instantaneous calculation mode Simulation benchmark: The real-time measured data from field sensors is used as the comparison benchmark; Time step: Millisecond-level calculation step; Convergence criterion: MAPE of simulation results and real-time measured data ≤ 2.0%; Operation process: Input real-time sensor data, perform millisecond-level calculations through component-level polynomial fitting surrogate model, correct simulation results using real-time error compensation module, and output timing parameters such as real-time total power consumption and integrated COP of the system to accurately reflect the current operating status of the system.

[0030] (2) Unsteady-state simulation mode Input boundaries: initial parameters for system cold start, boundary conditions for load switching across zones (such as load threshold for low load → medium load). Output field list: Equipment status smooth transition curve, startup inrush current data (peak current, duration), energy consumption dynamic change curve, chiller unit COP dynamic change curve, water pump power dynamic change curve; Operation process: Simulate the entire process of the system from cold start to stable operation, as well as the transitional conditions of load switching across zones, and output the above field data to provide data support for start-stop control and transitional condition optimization.

[0031] III. Multi-objective optimization design Optimization algorithm: NSGA-II non-dominated sorting genetic algorithm to find the Pareto optimal solution for multiple objectives; Optimization objectives: Four objectives are optimized in a coordinated manner, and all objectives are solved based on time series values ​​for the next 1-4 hours; Objective 1: Minimize total system energy consumption; Objective 2: Maximize overall system COP; Objective 3: Minimize peak-valley load difference of the power grid; Objective 4: Minimize energy consumption fluctuation under transition conditions.

[0032] Optimization variables: chiller start / stop combination, chilled water supply temperature setpoint (7℃-12℃), water pump operating frequency (30Hz-60Hz) and variable frequency / fixed frequency mode, number of cooling tower fans (1-8 units) and spray intensity (80-150L / m²). 2 ·h); Constraints: 100% satisfaction of the cooling load predicted by the LSTM model for the next 1-4 hours, and equipment operating parameters within the rated operating range; Adaptive weight adjustment: The weights of each objective in multi-objective optimization are adaptively adjusted according to the load change rate; when the load change rate is at a low level, the weights of the objectives of minimizing energy consumption and maximizing overall energy efficiency are increased; when the load change rate is at a high level, the weights of the objective of smooth transition process are increased.

[0033] Step 4: Optimal control command generation and issuance The target operating parameters of the HVAC system are determined based on the multi-objective optimization results, and control commands are generated based on the target operating parameters. The control commands are then sent to the control layer (DDC) of the HVAC system for execution. Specifically, a massive "virtual trial run" is conducted through a digital twin to generate a Pareto optimal solution set. A compromise solution is selected from the Pareto solution set and sent as a sequence of feedforward control commands for the next 1-4 hours.

[0034] Step 5: Adaptive Model Update When the preset update trigger condition is met, the Long Short-Term Memory Neural Network (LSTM) prediction model and the component-level polynomial fitting surrogate model are incrementally updated based on the recently collected measured operating data to improve the matching degree between the prediction results and simulation results and the actual operating state.

[0035] The triggering conditions for the adaptive update of the model include at least one of the following: update is triggered after equipment maintenance, update is triggered when the prediction error exceeds a preset threshold within a consecutive preset time period, and update is triggered when the load interval distribution characteristics change beyond a preset threshold. In this invention, the specific triggering conditions are: after equipment maintenance, the average prediction error exceeds 3.5% for 30 consecutive days, and the load interval proportion changes by more than 20%. The average prediction error is the average absolute percentage error between the predicted and actual values ​​of energy consumption and energy efficiency ratio for the next 4 hours.

[0036] The incremental model update process is as follows: extract the actual operating data of the HVAC system from the past 30 days, and calculate the correction amount of the component-level polynomial fitting model coefficients and the correction amount of the LSTM model weights after preprocessing; update the model parameters through incremental training to avoid overfitting and increased training costs caused by full training; perform accuracy verification after the update, and the update will take effect if the verification is successful. At the same time, store the historical version of the model. If the model performance does not improve after the update, immediately roll back to the optimal model version before the update to ensure the continuity and reliability of the system optimization control.

[0037] (iv) Beneficial effects By introducing a high-order polynomial fitting surrogate model with a coupling correction term and collaborating with a bidirectional LSTM time-series prediction model with an attention mechanism, high-fidelity simulation (error ≤ 2.0%) and accurate prediction of energy consumption / efficiency in the next 1-4 hours (error ≤ 3.2%) of HVAC systems are achieved, providing a reliable foundation for optimized control.

[0038] A complete closed loop of "prediction-simulation-optimization-execution-update" is constructed, and the NSGA-II algorithm is used for online multi-objective optimization. The weights can be adaptively adjusted according to load changes, simultaneously reducing the total energy consumption of the system, improving the overall COP, smoothing the fluctuations of grid load and transitional operating conditions, and achieving optimal energy efficiency in dynamic environments while meeting the predicted cooling load demand.

[0039] Through the deviation-triggered incremental update and rollback mechanism of the model, it effectively copes with equipment aging and changes in operating scenarios, ensuring the long-term stability of optimization effects, adapting to all working conditions and complex operating scenarios, and requiring no hardware modification, with strong engineering adaptability. Attached Figure Description

[0040] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0041] (a) Implementation prerequisites and system configuration (1) Hardware environment In this embodiment, the core equipment configuration of the HVAC system includes: 2 chillers with a rated cooling capacity of 900RT, 1 chiller with a rated cooling capacity of 400RT, 4 chilled water pumps (2 variable frequency + 2 fixed frequency), 4 cooling water pumps (2 variable frequency + 2 fixed frequency), and 8 cooling towers.

[0042] System hardware requirements: CPU is Intel Core i5-7400T / 2.40GHz or higher, memory is 8GB or higher, and hard disk is 250GB or higher; equipped with an edge computing gateway to support real-time preprocessing of sensor data and local deployment of model inference with latency ≤50ms; equipped with data acquisition devices such as temperature sensors, flow sensors, power sensors, wind speed sensors, voltage and current phase sensors, etc.

[0043] (2) Software environment Operating System: Windows 7 or later 64-bit system; Development and Run Platform: MATLAB and Python-TensorFlow parallel computing framework (MATLAB is used for component physics modeling and simulation, and Python-TensorFlow is used for fast training of LSTM models, improving training efficiency by 30%); Database: A database with time series data partitioning storage function, classified and stored according to load level (low / medium / high / ultra-high), improving query speed by 40%, used to store measured data, model parameters and calculation results.

[0044] (3) Data acquisition and preprocessing The data collection period was 18 months, acquiring a total of 13,140 sets of hourly data. The data covered all dimensions of equipment operating status data, environmental data, and load-related data, specifically historical operating parameters (cooling load, unit start-up / shutdown status, pump / fan frequency, etc.), external environmental data (outdoor dry-bulb / wet-bulb temperature, solar radiation intensity, wind speed, etc.), and target variables (total system power consumption, COP, power consumption of individual devices, etc.). Outliers were removed using a combination of the 3σ criterion and the isolated forest algorithm. Maximum-minimum normalization was used to map the data to the [0,1] interval. The standardization formula is as follows: Data allocation: 60% for LSTM model training, 25% for model validation, and 15% for model testing. Cross-validation is introduced to avoid overfitting.

[0045] (II) Construction and Implementation of Component-Level Polynomial Fitting Proxy Model 1. Chiller unit modeling COP data were collected from two 900RT and one 400RT chiller units under PLR (10%-100%) and cooling water inlet temperature (18℃-30℃). Polynomial fitting coefficients were calculated using a "piecewise fitting + global calibration" strategy. Determine the equipment aging coefficient A fourth-order bivariate polynomial fitting model including equipment aging correction terms was constructed, and the model determination coefficient was... For example, when the cooling water inlet temperature of the 900RT chiller is 18℃ and the PLR ​​is 100%, the model output COP is 9.532, which is consistent with the measured value; when the cooling water inlet temperature of the 400RT chiller is 18℃ and the PLR ​​is 80%, the model output COP is 9.647, which is consistent with the measured value.

[0046] 2. Water pump modeling For variable frequency / fixed frequency water pumps, measured data on head, power, efficiency, and pipeline pressure loss across the entire flow range were collected. A fourth-order univariate polynomial model including a coupling term for pipeline pressure loss was constructed. Taking a 900RT chiller unit equipped with a variable frequency cooling water pump as an example, the flow-head model coefficients were fitted to obtain the model parameters. Pressure loss coefficient ,Model At an operating frequency of 50Hz and a pipeline pressure loss of 18.7kPa, the model output head was 30.1m and the power was 60.4kW, which was consistent with the measured values.

[0047] 3. Cooling Tower Modeling Based on the energy conservation and enthalpy-humidity diagram calculation method, a thermodynamic model of the cooling tower is constructed, and the wind speed correction coefficient (0.02), spray intensity coupling coefficient (0.05), and spray pump power consumption coefficient are determined. Example calculation: Input cooling water flow rate 3200 m³ / h 3 / min, cooling water inlet temperature 30℃, outlet temperature 27℃, outdoor wet bulb temperature 27.6℃, wind speed 2.3m / s, spray intensity 100L / m 2 •h, model output airflow 2580m³ 3 / min, fan power consumption 13.86kW, spray pump power consumption 5.0kW, error from actual value ≤1.5%.

[0048] (III) Implementation process of LSTM-based energy consumption time series prediction The preprocessed 13,140 sets of data were input into a bidirectional LSTM prediction model with attention mechanism. The multidimensional running data of the previous 24 hours were used as input, and the training / validation / test sets were divided into 6:2.5:1.5. The Adam optimizer was used for training, and the parameters were optimized by combining a dynamic learning rate adjustment strategy and an early stopping mechanism. The initial learning rate of the Adam optimizer was set to 0.001, batch size=32, and 100 iterations were performed. The early stopping mechanism was introduced.

[0049] After training, the test set validation results showed that the MAPE for the predicted energy consumption in the next 4 hours was 3.1%, and the MAPE for the predicted comprehensive COP was 2.9%, both meeting the accuracy requirements. The MAPE for the predicted energy consumption in summer was 3.3%, and in winter it was 2.9%. The MAPE for the predicted COP in summer was 3.1%, and in winter it was 2.7%, both exceeding the accuracy targets. The model ultimately outputs the time-series prediction sequences of the total system energy consumption and comprehensive COP for the next 1-4 hours, providing driving input for digital twin simulation.

[0050] Based on the load sequence predicted by LSTM, a mapping model between the load change rate and the optimal state of the equipment is constructed to achieve accurate matching between dynamic load and equipment state. For example, when the load change rate is 2% / h (stable), it matches the state of the chiller unit running at full load and the water pump running at high frequency; when the load change rate is 20% / h (drastic change), it matches the state of the chiller unit gradually increasing the load and the water pump frequency slowly increasing.

[0051] (iv) Implementation process of multi-objective optimization driven by digital twin 1. Dual-modal simulation operation Instantaneous calculation mode: Input the measured cooling load of 1800RT, cooling water inlet temperature of 28℃, and outdoor wet bulb temperature of 27.6℃ at a certain moment. The model outputs the total power consumption of the system in milliseconds as 1268.5kW, with a comprehensive COP of 5.32. The measured value MAPE is 1.8%≤2.0%, which meets the convergence criterion. Unsteady-state simulation mode: Simulates the cold start-up process when the cooling load increases from 150RT to 450RT, outputting the peak starting inrush current of 187A (lasting 0.3s) and the transition curve of the chiller unit COP increasing from 2.13 to 5.67, providing data support for the optimization of transition conditions.

[0052] 2. Multi-objective optimization and instruction execution Using the LSTM-predicted cooling load of 1800RT (ultra-high load) for the next 4 hours as input, the NSGA-II algorithm is used for multi-objective optimization. The optimization objectives are to minimize the total system energy consumption, maximize the overall system COP, minimize the peak-valley load difference of the power grid, and minimize the energy consumption fluctuation under transition conditions. The optimization variables are the chiller start-stop combination, the chilled water supply temperature setpoint, the pump operating frequency and variable frequency / fixed frequency mode, the number of cooling tower fans and the spray intensity. The constraint is to meet 100% of the predicted cooling load demand.

[0053] At this point, the load change rate is 3% / h (stable). The target weights are adaptively adjusted based on the load change rate, with the weights allocated as follows: energy consumption minimization 0.4, COP maximization 0.4, peak-valley load difference minimization 0.2, and transition smoothness 0. A Pareto optimal solution set is generated through virtual trial operation using a digital twin. A compromise solution is selected from the Pareto solution set as the optimal control scheme.

[0054] (v) Generation and Implementation of Optimal Control Commands The compromise solution selected from the Pareto optimal solution set is as follows: Start two 900RT chiller units, with the chilled water pumps (two variable frequency drives) operating at a frequency of 52Hz, the cooling water pumps (two variable frequency drives) operating at a frequency of 57Hz, and start six cooling tower fans with a spray intensity of 120L / m². 2 h. Based on this compromise solution, control commands are generated and sent to the DDC control layer of the HVAC system for execution.

[0055] After the instruction was issued, the total energy consumption of the system measured on site was 1268.5kW, the comprehensive COP was 5.32, the peak-valley load difference of the power grid was reduced by 25%, and the energy consumption fluctuation under the transition condition was 7.5%≤8%. Under the premise of meeting the predicted cooling load demand of 1800RT, the multi-objective optimization objectives were achieved.

[0056] (vi) Implementation process of adaptive model update In this embodiment, the update triggering conditions set in the claims are strictly followed: after equipment maintenance, the average prediction error exceeds 3.5% for 30 consecutive days, and the load range ratio changes by more than 20%. The average prediction error is the average absolute percentage error between the predicted and actual values ​​of energy consumption and energy efficiency ratio for the next 4 hours.

[0057] After the chiller unit completes heat exchanger cleaning, an incremental model update is triggered: 30 days of measured data after cleaning are extracted, and the LSTM prediction model and component-level polynomial fitting surrogate model are incrementally updated. The LSTM model weights and chiller unit polynomial coefficients are updated during incremental training. Validation set testing demonstrates the model's performance. =0.9996≥0.9993, prediction error MAPE=3.0%≤3.5%, update takes effect; if the updated validation set If 0.9990 < 0.9993, then immediately roll back to the model version before cleaning to ensure the continuity of system operation.

[0058] (vii) Verification of the implementation effect under all working conditions The method of this invention is applied to the above-mentioned HVAC system, covering all operating conditions from low load (300RT), medium load (650RT), high load (1200RT), ultra-high load (1800RT), to extreme load (2100RT). Under each operating condition, multi-objective optimization and optimal energy efficiency are achieved while meeting the predicted cooling load demand. For example: Low load 300RT: Prioritize the use of small-capacity 400RT units, increasing the overall system COP from 2.50 to 4.83, an increase of 93.2%; Ultimate load 2100RT: With three units operating in tandem and optimized load ratio, the system's overall COP reaches 4.94, achieving optimal energy efficiency under ultimate load.

[0059] Overall Implementation Effectiveness Verification: Component Model The system simulation error MAPE is ≤2.0%, the prediction error MAPE for the next 4 hours is ≤3.2%, the annual comprehensive COP is increased from 3.2 to 5.38, the annual energy saving rate is 42.5%, the peak-valley load difference of the power grid is reduced by 25%, and the energy consumption fluctuation under the transition condition is ≤8%. The invention fully achieves its objectives and requires no hardware modification, with strong engineering adaptability.

[0060] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A method for predicting HVAC digital twin energy efficiency based on LSTM and polynomial fitting, characterized in that, Includes the following steps: S1 Data Acquisition and Preprocessing: Collect historical and real-time operating data of the HVAC system, and perform outlier removal and standardization preprocessing on the data to obtain the input dataset for modeling and prediction; the operating data includes at least equipment operating status data, environmental data, and load-related data; S2 Energy Consumption Time Series Prediction Based on LSTM: A Long Short-Term Memory Neural Network (LSTM) prediction model is constructed based on preprocessed time series operating data. Using multi-dimensional operating data within a preset historical time window as input, the model performs rolling predictions on the total energy consumption and comprehensive energy efficiency ratio (COP) of the HVAC system within a preset prediction time domain, and outputs a time series prediction sequence of energy consumption and energy efficiency indicators. S3 Digital Twin Dynamic Simulation and Multi-Objective Optimization: The future cooling load demand predicted by S2 is used as the driving input and input into the HVAC digital twin constructed based on the component-level polynomial fitting surrogate model to dynamically simulate the future operating state of the system. The multi-objective optimization algorithm is used to solve the system operating parameters to obtain an optimized control scheme that meets the preset constraints. The objectives of the multi-objective optimization include at least: minimizing the total system energy consumption, maximizing the overall system COP, minimizing the peak-valley load difference of the power grid, and minimizing energy consumption fluctuations during transitional operating conditions. The optimization variables include: chiller start-stop combination, chilled water supply temperature setpoint, water pump operating frequency and variable frequency / fixed frequency mode, number of cooling tower fans and spray intensity. The preset constraints include: meeting the predicted cooling load demand; S4 Optimal Control Command Generation and Issuance: Determine the target operating parameters of the HVAC system based on the multi-objective optimization results, generate control commands based on the target operating parameters, and issue the control commands to the control layer of the HVAC system for execution; S5 Model Adaptive Update: When the preset update trigger condition is met, the Long Short-Term Memory Neural Network (LSTM) prediction model and the component-level multinomial fitting surrogate model are incrementally updated based on the recently collected measured running data to improve the matching degree between the prediction results and simulation results and the actual running state.

2. The method according to claim 1, characterized in that, In step S1, the operational data includes historical operational parameters, external environment data, and target variables; The historical operating parameters include cooling load, chiller start / stop status, pump and fan operating frequency, chilled water supply and return water temperature, cooling water inlet and return water temperature, pump operating mode, and pipeline pressure loss data collected at preset time intervals. The external environmental data includes outdoor dry-bulb temperature, wet-bulb temperature, solar radiation intensity, and wind speed collected at preset time intervals; The target variables include total system power consumption, overall energy efficiency ratio (COP), chiller power consumption, water pump power consumption, and cooling tower fan power consumption.

3. The method according to claim 1, characterized in that, In step S1, outlier removal employs a combination of the 3σ criterion and the isolated forest algorithm. Standardization uses a maximum-minimum normalization method to map the original data to a preset interval. The standardization formula is: in, For the original data, and These represent the minimum and maximum values ​​of the data, respectively.

4. The method according to claim 1, characterized in that, In step S2, the hidden layer of the Long Short-Term Memory (LSTM) neural network prediction model adopts a multi-layer bidirectional LSTM structure and includes an attention mechanism layer; the output layer is used to output the time-series prediction results of the system's total energy consumption and overall energy efficiency ratio (COP) within a preset future time domain. The LSTM prediction model is trained using the Adam optimizer and optimized for parameters by combining a dynamic learning rate adjustment strategy and an early stopping mechanism.

5. The method according to claim 1, characterized in that, In step S3, the component-level polynomial fitting surrogate model driving the digital twin simulation includes: a) Chiller unit model: A polynomial fitting model including equipment aging correction terms is adopted. Taking partial load rate and cooling water inlet temperature as inputs, the output is the chiller unit's comprehensive energy efficiency parameter, expressed as: in, For partial load factor, This refers to the inlet temperature of the cooling water. These are the polynomial fitting coefficients. This refers to the equipment aging factor. The model's coefficient of determination is the cumulative runtime of the equipment. ; b) Pump Model: A polynomial fitting model including pipeline pressure loss coupling terms is adopted. The pump operating frequency is used as input, and the output parameters include pump head, power, and efficiency, including: Flow-head model: , Flow-power model: , Flow-efficiency model: , and frequency conversion switching loss model: , in, For the first The operating frequency of the water pump For pipeline pressure loss, , The coefficients of determination for each model are selected based on the coefficients obtained by fitting the measured data. c) Cooling tower model: A thermodynamic model constructed based on the principle of energy conservation and the calculation method of air enthalpy-humidity diagram. Input variables include cooling water flow rate, inlet and outlet water temperature, outdoor wet-bulb temperature, wind speed, and spray intensity. Through enthalpy calculation, mass flow rate ratio calculation, and iterative verification, the outputs are air flow rate, fan power consumption, and spray pump power consumption.

6. The method according to claim 1, characterized in that, In step S3, the weights of each objective in the multi-objective optimization are adaptively adjusted according to the load change rate; when the load change rate is at a low level, the weights of the objectives of minimizing energy consumption and maximizing overall energy efficiency are increased; when the load change rate is at a high level, the weight of the objective of smoothness of transition process is increased.

7. The method according to claim 1, characterized in that, In step S3, the digital twin simulation operation includes two modes: Instantaneous computing mode: used to input real-time sensor data and quickly output the system's real-time total power consumption and overall energy efficiency ratio (COP) based on a component-level proxy model; Unsteady-state simulation mode: Used to simulate the dynamic response of an HVAC system from startup to stable operation, and output parameters of the equipment state transition process.

8. The method according to claim 1, characterized in that, In step S5, the triggering conditions for the adaptive update of the model include at least one of the following: update is triggered after equipment maintenance, update is triggered when the prediction error exceeds a preset threshold within a continuous preset time period, and update is triggered when the change in load interval distribution characteristics exceeds a preset threshold. The triggering conditions for the adaptive update of the model are: after equipment maintenance, the average prediction error exceeds 3.5% for 30 consecutive days, and the change in the proportion of the load range exceeds 20%. The mean of the prediction error is the average absolute percentage error between the predicted and actual values ​​of energy consumption and energy efficiency ratio for the next 4 hours.

9. The method according to claim 1, characterized in that, The method, while meeting the predicted cooling load demand, achieves improved overall energy efficiency of HVAC systems, reduced total system energy consumption, reduced grid-side load fluctuations, and improved smoothness of transitional operating conditions.

10. The method according to claim 1, characterized in that, The method is applicable to central air conditioning systems that include refrigeration units, chilled water pumps, cooling water pumps, and cooling towers, and is suitable for system operation optimization control under different load conditions.

Images