Intelligent predictive regulation and control system and method for multi-energy complementary electric floor heating

Abstract

The invention provides a multi-energy complementary electric floor heating intelligent prediction regulation and control system and method. Comprising a sensing module, a prediction module, an optimization regulation and control module, a user feedback module and the like, and efficient cooperative operation of photovoltaic, energy storage, electric floor heating and a power grid under the sub-plateau complex climate condition is achieved by constructing a closed-loop intelligent regulation and control system of'sensing-prediction-optimization-feedback '. High-resolution data acquisition is realized by adopting an Internet of Things sensing network, and dominant features are screened by applying a random forest model, so that the effectiveness of input data and the model training efficiency are remarkably improved. According to long-term verification of actual farm houses, compared with a traditional control mode, heating comfort is guaranteed, meanwhile, system energy consumption and operation cost are remarkably reduced, the utilization rate of renewable energy sources is increased, good operation stability is kept under extreme weather, and energy conservation and emission reduction are achieved. And a set of efficient, reliable and generalizable intelligent solution is provided for clean heating of the sub-plateau and similar cold regions.

Description

Technical Field

[0001] This invention relates to the field of building energy management technology, and in particular to a multi-energy complementary electric underfloor heating intelligent prediction and control system and method. Background Technology

[0002] Currently, under the dual-carbon goals, clean and intelligent building heating has become an important direction for energy transition. Sub-plateau regions such as Dingxi in Gansu and Haidong in Qinghai are characterized by high altitude, strong radiation, large diurnal temperature variations, and persistently low winter temperatures. Traditional rural housing heating relies heavily on direct heating methods such as electric heaters or simple electric underfloor heating, resulting in high energy consumption, slow response, and poor comfort. In recent years, multi-energy complementary systems consisting of photovoltaics, energy storage, and electric underfloor heating have been gradually applied to rural residences, which can utilize renewable energy to some extent and reduce dependence on the power grid. However, existing systems still face many challenges in actual operation: First, the complex meteorological conditions in sub-plateau regions significantly affect photovoltaic output due to the coupling effect of radiation and module temperature, causing drastic fluctuations; second, the strong thermal inertia of buildings means that electric underfloor heating systems exhibit significant lag characteristics, making it difficult for traditional control strategies to achieve timely responses; third, the high uncertainty of user behavior makes it difficult to meet personalized comfort needs with fixed temperature settings; and fourth, existing control methods are mostly based on rules or classic PID control, which cannot achieve coordinated optimization of multiple objectives such as energy saving, comfort, and economy under multiple constraints. In addition, most studies remain at the simulation stage and lack full-link verification in real building environments, resulting in insufficient generalization ability and reliability of the models in real-world scenarios. Summary of the Invention

[0003] In view of this, in order to solve the problems existing in the technical background, the present invention proposes a multi-energy complementary electric underfloor heating intelligent predictive control system and method. Specifically, it includes the following:

[0004] A multi-energy complementary electric underfloor heating intelligent predictive control system, characterized in that it includes: a sensing module, a predictive module, an optimization control module, and a user feedback module;

[0005] The sensing module is used to collect meteorological data, photovoltaic power generation data, energy storage system data, electric floor heating system data, indoor temperature data, and user settings and behavior data to construct a multi-source high-resolution dataset.

[0006] The prediction module includes a feature selection unit, a photovoltaic output prediction unit, and a load multi-scale prediction unit. The feature selection unit identifies the dominant features affecting system operation based on a random forest model. The photovoltaic output prediction unit predicts photovoltaic output based on a decoupled model of photovoltaic module temperature and irradiance. The load multi-scale prediction unit uses a spatiotemporal graph attention network model to integrate building spatial topology and time series to achieve multi-step prediction of indoor temperature and load. The optimization and control module uses an improved reinforcement learning algorithm, with the output of the prediction module as input, to construct a multi-objective reward function that includes energy consumption, comfort, system stability, and economy, and outputs continuous control commands for electric underfloor heating power, energy storage system charging and discharging power, and grid interaction power.

[0007] The user feedback module is used to collect user comfort evaluation and behavior data, build a user preference model, and dynamically integrate user feedback information into the reward function of the optimization and control module to achieve adaptive adjustment of the heating strategy.

[0008] In one embodiment of the present invention, the sensing module includes an Internet of Things (IoT) sensor network deployed inside and outside the building and on the side of energy equipment. The sensor network collects data at a resolution of not less than 1 minute and performs data cleaning, alignment and storage through edge computing and cloud platform.

[0009] In one embodiment of the present invention, the decoupling model used by the photovoltaic power output prediction unit is as follows:

[0010] ;

[0011] in: ,but = ;like ,but This indicator can effectively compensate for the decrease in efficiency caused by temperature rise, making the irradiation input more in line with physical laws.

[0012] In one embodiment of the present invention, the load multi-scale prediction unit uses a spatiotemporal graph attention network model to construct a spatial graph with each room in the building as a node. It dynamically learns the thermal coupling weights between rooms through a graph attention mechanism, and uses a temporal network combining dilated causal convolution and gated recurrent units to capture the multi-scale temporal dynamics of the system, and outputs a predicted sequence of indoor temperature, electric underfloor heating power demand and photovoltaic available energy for each room in the next 15 to 240 minutes.

[0013] In one embodiment of the present invention, the improved reinforcement learning algorithm used by the optimization and control module is a dual-delay deep deterministic strategy gradient algorithm based on priority experience replay. Its state space includes the current indoor temperature, outdoor temperature, wind speed, effective irradiance, energy storage state of charge, electric underfloor heating power, user-set temperature, user feedback information, and future multi-step prediction sequence output by the load multi-scale prediction unit; its action space is a continuous value control command for the electric underfloor heating set power, energy storage charging power, and energy storage discharging power.

[0014] In one embodiment of the present invention, the multi-objective reward function of the optimization and control module is composed of the following weighted summation:

[0015] ;

[0016] in These are weighting coefficients, which can be optimized through experiments.

[0017] In one embodiment of the present invention, the user preference model of the user feedback module is constructed based on a Gaussian temperature utility function:

[0018] ;

[0019] Where: T: current indoor temperature; Tpref: user preferred temperature (dynamically changing); σ: temperature tolerance.

[0020] A method for intelligent prediction and control of multi-energy complementary electric underfloor heating, characterized in that it is applied to the intelligent prediction and control system of multi-energy complementary electric underfloor heating as described in any one of claims 1-7, and includes the following steps:

[0021] S1: Continuously collect multi-source operational data through the perception module, and perform data cleaning and preprocessing to build a high-quality training and inference dataset;

[0022] S2: Use the random forest model to perform feature importance analysis on the dataset and select the dominant features;

[0023] S3: Based on the decoupled model of photovoltaic module temperature and irradiance, predict photovoltaic output;

[0024] S4: Based on the spatiotemporal graph attention network model, it integrates building space topology and historical time series data to achieve prediction of indoor temperature and system load in multiple future steps;

[0025] S5: Based on user comfort evaluation and behavioral data, dynamically update the preference temperature and tolerance parameters in the user preference model;

[0026] S6: Using the features selected in step S2, the prediction results of steps S3 and S4, and the user preference information in step S5 as state inputs, the reinforcement learning optimization control model based on priority experience replay is used to calculate and output the optimal electric underfloor heating power, energy storage charging and discharging power, and grid interaction power control commands.

[0027] S7: Executes control commands and continuously updates the prediction model and optimizes the control strategy based on system operation results and user feedback.

[0028] The above technical solution has the following beneficial effects:

[0029] This invention provides a multi-energy complementary electric underfloor heating intelligent prediction and control system and method. By constructing a closed-loop intelligent control system of "sensing-prediction-optimization-feedback," it achieves efficient collaborative operation of photovoltaics, energy storage, electric underfloor heating, and the power grid under complex sub-plateau climate conditions. High-resolution data acquisition is achieved using an Internet of Things (IoT) sensor network, and a random forest model is used to select dominant features, significantly improving the effectiveness of input data and model training efficiency. The introduction of a photovoltaic temperature-irradiance decoupling model effectively compensates for the impact of component temperature rise on power generation efficiency, substantially improving the accuracy of photovoltaic output prediction. A spatiotemporal graph attention network model, through dynamic learning of the thermal coupling relationship and time lag characteristics of building space, achieves high-precision prediction of future multi-step loads and room temperature, providing reliable forward-looking information for optimized control. This invention has been verified through long-term testing in actual rural residences. Compared with traditional control methods, it significantly reduces system energy consumption and operating costs while ensuring heating comfort, improves the utilization rate of renewable energy, and maintains good operational stability under extreme weather conditions. It provides an efficient, reliable, and scalable intelligent solution for clean heating in sub-plateau and similar cold regions. Attached Figure Description

[0030] Figure 1 The experimental building BIM model diagram of this invention Detailed Implementation

[0031] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Example 1: A multi-energy complementary electric underfloor heating intelligent predictive control system, characterized in that it includes: a sensing module, a predictive module, an optimization control module, and a user feedback module;

[0033] The sensing module is used to collect meteorological data, photovoltaic power generation data, energy storage system data, electric floor heating system data, indoor temperature data, and user settings and behavior data to construct a multi-source high-resolution dataset.

[0034] The prediction module includes a feature selection unit, a photovoltaic output prediction unit, and a load multi-scale prediction unit. The feature selection unit identifies the dominant features affecting system operation based on a random forest model. The photovoltaic output prediction unit predicts photovoltaic output based on a decoupled model of photovoltaic module temperature and irradiance. The load multi-scale prediction unit uses a spatiotemporal graph attention network model to integrate building spatial topology and time series to achieve multi-step prediction of indoor temperature and load. The optimization and control module uses an improved reinforcement learning algorithm, with the output of the prediction module as input, to construct a multi-objective reward function that includes energy consumption, comfort, system stability, and economy, and outputs continuous control commands for electric underfloor heating power, energy storage system charging and discharging power, and grid interaction power.

[0035] The user feedback module is used to collect user comfort evaluation and behavior data, build a user preference model, and dynamically integrate user feedback information into the reward function of the optimization and control module to achieve adaptive adjustment of the heating strategy.

[0036] Example 2, based on Example 1, in this example, the sensing module includes an Internet of Things (IoT) sensor network deployed inside and outside the building and on the side of energy equipment. The sensor network collects data at a resolution of not less than 1 minute and performs data cleaning, alignment and storage through edge computing and cloud platform.

[0037] The decoupling model used in the photovoltaic power output prediction unit is as follows:

[0038] ;

[0039] in: ,but = ;like ,but This indicator can effectively compensate for the decrease in efficiency caused by temperature rise, making the irradiation input more in line with physical laws.

[0040] The load multi-scale prediction unit uses a spatiotemporal graph attention network model to construct a spatial graph with each room in the building as a node. It dynamically learns the thermal coupling weights between rooms through a graph attention mechanism and uses a temporal network combining dilated causal convolution and gated recurrent units to capture the multi-scale temporal dynamics of the system. It outputs the predicted sequence of indoor temperature, electric underfloor heating power demand and photovoltaic available energy for each room in the next 15 to 240 minutes.

[0041] Example 3, based on Example 1, uses an improved reinforcement learning algorithm in the optimization and control module as a dual-delay deep deterministic strategy gradient algorithm based on priority experience replay. Its state space includes the current indoor temperature, outdoor temperature, wind speed, effective irradiance, energy storage state of charge, electric underfloor heating power, user-set temperature, user feedback information, and future multi-step prediction sequence output by the load multi-scale prediction unit; its action space is a continuous value control command for the electric underfloor heating set power, energy storage charging power, and energy storage discharging power.

[0042] In one embodiment of the present invention, the multi-objective reward function of the optimization and control module is composed of the following weighted summation:

[0043] ;

[0044] in These are weighting coefficients, which can be optimized through experiments.

[0045] The user preference model of the user feedback module is constructed based on a Gaussian temperature utility function.

[0046] ;

[0047] Where: T: current indoor temperature; Tpref: user preferred temperature (dynamically changing); σ: temperature tolerance.

[0048] A method for intelligent prediction and control of multi-energy complementary electric underfloor heating, characterized in that it is applied to the intelligent prediction and control system of multi-energy complementary electric underfloor heating as described in any one of claims 1-7, and includes the following steps:

[0049] S1: Continuously collect multi-source operational data through the perception module, and perform data cleaning and preprocessing to build a high-quality training and inference dataset;

[0050] S2: Use the random forest model to perform feature importance analysis on the dataset and select the dominant features;

[0051] S3: Based on the decoupled model of photovoltaic module temperature and irradiance, predict photovoltaic output;

[0052] S4: Based on the spatiotemporal graph attention network model, it integrates building space topology and historical time series data to achieve prediction of indoor temperature and system load in multiple future steps;

[0053] S5: Based on user comfort evaluation and behavioral data, dynamically update the preference temperature and tolerance parameters in the user preference model;

[0054] S6: Using the features selected in step S2, the prediction results of steps S3 and S4, and the user preference information in step S5 as state inputs, the reinforcement learning optimization control model based on priority experience replay is used to calculate and output the optimal electric underfloor heating power, energy storage charging and discharging power, and grid interaction power control commands.

[0055] S7: Executes control commands and continuously updates the prediction model and optimizes the control strategy based on system operation results and user feedback.

[0056] This embodiment is based on a field measurement and experimental platform established in a typical rural house in Dingxi, Gansu Province. It constructs a full-link data acquisition and energy flow monitoring system covering photovoltaics, energy storage, electric underfloor heating, and indoor thermal environment, and forms a unified description system for BIM, IoT, and multi-source energy flow, providing reliable data support for subsequent AI prediction models and reinforcement learning control algorithms. This embodiment includes regional climate characteristics, rural house building characteristics, heating system structure, test system composition, data acquisition methods, and experimental data characteristic analysis.

[0057] To verify the applicability of the proposed AI-MEC electric underfloor heating system in typical cold sub-plateau regions, Dingxi, Gansu Province, was selected as the experimental area. Dingxi City is located in the transition zone between the Loess Plateau and the Qinghai-Tibet Plateau in northwestern China (35.58°N, 104.62°E), at an altitude of approximately 1600m, and belongs to a typical cold and dry sub-plateau climate. It is characterized by low winter temperatures, large diurnal temperature variations, strong radiation, and high wind speeds, making it a typical representative area for multi-energy complementary systems in cold regions.

[0058] The extreme minimum temperature in winter can reach -15°C, with a seasonal average temperature of -6.5°C (standard deviation ±3.2°C). The diurnal temperature range is typically between 15 and 20°C. The average wind speed is approximately 3.5 m / s (peak value 8 m / s, standard deviation ±1.8 m / s). Relative humidity remains between 38% and 48%, with an average of 42%. Atmospheric pressure is approximately 85.2 kPa. The peak total solar irradiance is approximately 820 W / m². 2 The average daily radiation is 10.8 MJ / m². 2 The average annual solar radiation is 38.5°C, with direct radiation accounting for 62% and diffuse radiation accounting for 38%. The Clear Sky Index (CI) is approximately 0.63, indicating that winters are predominantly characterized by clear and cold weather. According to the climate zoning standards of the National Standard for Thermal Design of Civil Buildings (GB50176-2016), Dingxi belongs to Cold Region II, with an annual heating degree days (HDD). 18 The energy consumption is approximately 4200–4500 K·d, exhibiting typical sub-plateau heating energy consumption characteristics.

[0059] This invention, building BIM, such as Figure 1As shown, based on the aforementioned climatic characteristics of Dingxi area, this invention selects a typical rural residence in Lintao County, Dingxi City, Gansu Province as the experimental building. This building is located in a sub-plateau cold and dry climate zone at an altitude of approximately 1600 m, characterized by large diurnal temperature differences, strong radiation, and significant wind speed fluctuations in winter. The building is a single-story brick-concrete structure with a total floor area of ​​approximately 120 m². 2 The building has a floor height of 3.5m and a north-south orientation. The courtyard-style enclosure can effectively reduce the exposure rate of the windward side and the impact of wind pressure.

[0060] The equivalent heat transfer coefficients of various parts of the building envelope are shown in the table below. The heat transfer areas of the exterior walls, roof, ground, and doors and windows are 86.4 m². 2 120m 2 120m 2 and 22.5m 2 The volumetric heat capacity of each major component is 1.6–2.1 MJ / m³. 3 K, average density ρ = 2100 kg / m³ 3 The total building heat capacity is approximately 150 MJ / K. The design indoor temperature is set at 20°C (with an allowable fluctuation of ±1°C), and the air exchange rate is 0.5 times / h. Thermal inertia test results show that the indoor temperature response lags by approximately 2–3 hours (standard deviation ±0.4 hours), demonstrating good heat storage and buffering performance.

[0061]

[0062] This invention constructs a multi-energy complementary heating system centered on photovoltaic (PV), energy storage system (ESS), and electric underfloor heating (EFH), achieving multi-energy flow coupling and efficient utilization of renewable energy. It mainly comprises the following subsystems:

[0063] Photovoltaic (PV) Power Generation System: The experimental platform is equipped with an 8 kWp distributed rooftop photovoltaic power generation system to provide renewable energy support for rural residential heating systems and energy storage devices. The photovoltaic modules use high-efficiency monocrystalline silicon cells with a rated peak efficiency of 21.3%, and the inverter is a 5 kW single-phase grid-connected model with a maximum conversion efficiency of 97.5%. The system is installed at a tilt angle of 35°, facing due south, to adapt to the winter irradiance angle in Dingxi and improve energy capture capabilities under low solar altitude.

[0064] Due to the typical characteristics of strong irradiance and low temperature in Dingxi during winter, the output power of photovoltaic modules is affected by the irradiance intensity ( ) and component temperature ( The photovoltaic (PV) output exhibits a dual effect. In cold environments, the module temperature is typically in the low range of 5–15°C, resulting in high module efficiency. However, during the midday heat when irradiance rises rapidly, the module backsheet temperature can instantly surge to over 40°C, causing a decrease in PV conversion efficiency of approximately 6%–8%. Consequently, PV output displays a significant thermal-electric coupling effect and nonlinear dynamic response characteristics. Relying solely on irradiance cannot accurately predict PV output; therefore, module temperature must be introduced simultaneously, and a temperature-irradiance decoupling relationship must be established to provide the necessary data foundation for subsequent PV temperature model construction.

[0065] This invention configures a 10 kWh lithium iron phosphate (LiFePO4) energy storage system (ESS) as a key regulating unit between photovoltaic (PV) and electric underfloor heating systems. This system balances PV output fluctuations, smooths peak and valley loads, improves self-consumption rate, and enhances the overall operational stability of the multi-energy complementary system. The energy storage system constructs a complete energy flow and state monitoring system through a battery management system (BMS), inverter, current / voltage sensors, and control terminals, providing a realistic and quantifiable decision-making state space for subsequent reinforcement learning control algorithms. The ESS features intelligent charge and discharge management capabilities and can achieve dynamic load distribution during the winter heating season by operating the inverter in a bidirectional mode and forming a closed-loop coupling with the PV system.

[0066] This invention selects a cable-type electric underfloor heating system suitable for rural residential buildings as the main heat terminal of the heating system. The electric underfloor heating system features high thermal comfort, uniform temperature distribution, and low operating noise, and is particularly adaptable to the low outdoor temperatures, high wind speeds, and significant heat loss associated with traditional convection heating in sub-plateau regions. The experimental rural residence's electric underfloor heating system covers the living room and three bedrooms, with a total area of ​​approximately 80 m². 2 To ensure uniform system heating, a double-conductor heating cable winding method is used, and the cable spacing is adjusted according to the area of ​​each room to improve electrothermal conversion efficiency and thermal comfort level. The system's rated power is 120 W / m. 2 The total rated heating power is approximately 9.6 kW. In actual operation, on-demand heating is achieved through zoned control (independent circuits for the living room and bedroom), providing adjustable discrete units for enhanced learning-based regulation.

[0067] To achieve precise modeling and intelligent control of multi-energy complementary systems in rural residential areas in sub-plateau regions, this invention constructs a high-resolution Internet of Things (IoT) monitoring network covering meteorological, energy, building, and user sides. This system utilizes multi-source heterogeneous sensors to monitor all elements of photovoltaic (PV) power generation, energy storage systems (ESS), electric underfloor heating (EFH), indoor thermal environment, and user behavior, forming the core foundation for predictive models, reinforcement learning control strategies, and user feedback mechanisms. The entire system can achieve approximately 36,000 sets of multi-source data sampling per day, meeting the training requirements of ST-GAT load forecasting and PERTD3 control strategies. The testing system adopts a layered structure of "multi-point acquisition - edge computing - cloud storage - intelligent interface", consisting of the following four parts: The sensor layer is responsible for collecting meteorological, energy, environmental and user data to achieve high-frequency monitoring with a resolution of 1 minute; the edge processing layer is responsible for data preprocessing, fault monitoring and preliminary verification, including embedded controllers and local data buffer modules; the transmission network layer adopts a hybrid communication architecture of ZigBee + Wi-Fi + RS485 to ensure highly reliable transmission in rural environments; and the cloud platform layer connects to the database, AI model training platform and visualization interface to achieve unified data management and online processing.

[0068] In order to systematically evaluate the operational characteristics of multi-energy complementary electric underfloor heating systems in rural residential areas under complex meteorological conditions and to construct a reliable AI prediction and control model, this invention conducted continuous and systematic field measurements on the PV–ESS–EFH system throughout the winter.

[0069] The experiment lasted for more than 90 days (sampling every minute) and included the following typical scenarios: stable irradiance under clear weather conditions was defined as daily irradiance greater than 700 W / m². 2 Used for PV temperature model and EFH thermal inertia analysis; the low-irradiance condition under cloudy weather is determined as irradiance fluctuation less than 200 W / m. 2 It is used to analyze the ESS–EFH coordination characteristics under low-energy input conditions; the irradiation abrupt change condition (cloud cover) is defined as an irradiance decrease of more than 40% within 1–5 minutes, which is a key data source for training the feature decoupling model and RL control strategy; cold wave extreme low temperature condition Below -12°C, used to analyze peak building heat load and EFH operating limits; user behavior disturbance conditions, such as opening windows at night and users temporarily raising / lowering the set temperature, are used to train the user feedback mechanism of PERTD3. By covering multiple operating conditions, this invention constructs a realistic, complex, and highly nonlinear training environment, enabling the ST-GAT and PERTD3 models to have stronger generalization capabilities.

[0070] The present invention employs the following data cleaning process:

[0071] (1) Anomaly detection uses IQR (interquartile range) combined with continuous range method to identify abnormal jumps. PV output is automatically removed when it is greater than the theoretical upper limit. Wind speed change is manually verified when it is less than ±3 m / s·min.

[0072] (2) Missing data processing: if the missing time is less than 3 min, linear interpolation is used; if the missing time is 3–10 min, KNN interpolation is used; if the missing time is greater than 10 min, the time period is removed.

[0073] (3) Noise filtering: wavelet denoising is introduced to process high-frequency noise. The moving average window is 5 min. Savitzky-Golay filter is used to smooth the irradiance curve.

[0074] (4) Alignment and resampling: All data are unified to a 1 min resolution to solve the problem of sampling point time misalignment caused by communication delay.

[0075] (5) Data integrity check, check the synchronization of multi-dimensional features, and whether the significant drop in PV output is accompanied by Does the decrease in SOC accompany the discharge of ESS? Does the increase in EFH power accompany the increase in indoor temperature?

[0076] This invention establishes the following operating strategy: ESS safety strategy prohibits further discharge when SOC is less than 20%, prohibits further charging when SOC is greater than 90%, and prohibits charging at low temperatures (…). A current-limiting charging strategy is adopted (below −5°C). Over-temperature protection is provided in the EFH region. Automatic power-off when the temperature exceeds 33°C to prevent overheating and damage to flooring materials. Photovoltaic inverter protection: automatic disconnection for PV string overvoltage, undervoltage, and overcurrent, with nighttime backflow prevention protection. Communication anomaly fallback strategy: if IoT communication is interrupted for more than 10 minutes, the system automatically switches to traditional thermostat control. Extreme operating condition specifications: in extreme weather conditions such as cold waves and blizzards, priority is given to ensuring user safety and maintaining a basic indoor temperature (18°C).

[0077] This invention, based on over 90 days of continuous measured data, systematically analyzes the multidimensional data characteristics of meteorology, energy, indoor thermal environment, and user behavior. The test data exhibits significant non-steady-state characteristics, random perturbations, hysteresis coupling properties, and spatial correlation, providing crucial evidence for constructing AI prediction models and reinforcement learning control models.

[0078] Multi-energy complementary electric underfloor heating systems in rural residences in sub-plateau regions exhibit typical characteristics such as high disturbance, strong coupling, deep hysteresis, and user preference uncertainty, making it difficult for traditional control strategies to maintain stable and efficient operation in complex environments. Therefore, this invention, combining the aforementioned high-resolution measured data, constructs a "prediction-decision-feedback closed-loop intelligent control system" for real-world rural residential applications, realizing a complete model chain from data understanding, physical decoupling, trend prediction to real-time control and user adaptation.

[0079] Feature Understanding: Random Forest (RF) for Identifying Dominant Influencing Factors

[0080] Multi-energy complementary systems involve multi-dimensional variables such as PV, ESS, EFH, indoor temperature, and meteorological disturbances, resulting in a high-dimensional input space with complex correlations. Directly inputting all features into the prediction model and reinforcement learning environment leads to model redundancy, training instability, and slow convergence. Therefore, we first use Random Forest (RF) to perform feature importance analysis on 90 days of measured data: identifying the dominant factors affecting room temperature changes (such as outdoor temperature and wind speed); and extracting the key driving factors of PV output (…). The model eliminates redundant variables and reduces the dimensionality of the model input; it filters and trims the state spaces of ST-GAT and PERTD3; the RF module is equivalent to the "input cleaning and structured understanding layer", providing a theoretical basis for subsequent model construction.

[0081] Physical decoupling: Constructing PV temperature-irradiation ( The decoupling model shows that Dingxi region has strong sunlight and low temperature, but... Furthermore, the temperature and irradiation levels change rapidly, resulting in a significant "temperature-irradiation" coupled nonlinearity in PV output. Efficiency decreases as both irradiation and temperature increase; the temperature rise lags behind irradiation by approximately 10–20 minutes, causing PV output to not change synchronously with irradiation. Therefore, directly... Using it for prediction can lead to the accumulation of model errors. This invention constructs a physically constrained PV decoupled model, which corrects the influence of irradiance on power by adjusting component temperature, and generates more stable PV input characteristics: reducing noise propagation from PV fluctuations; improving the prediction stability of ST-GAT; avoiding RL errors in estimating available energy; this module is equivalent to a "physical prior layer", which preprocesses the most unstable energy source (PV) in the system.

[0082] Trend Prediction: The Spatiotemporal Graph Attention Network (ST-GAT) enables multi-step prediction of room temperature and energy flow. Electric underfloor heating systems exhibit significant thermal inertia and a lag characteristic of 30–50 minutes. Relying solely on real-time data for control can easily lead to under-adjustment / over-adjustment phenomena. Therefore, it is necessary to predict the future state of the system before regulation. The ST-GAT model uses: a graph structure to describe the spatial thermal coupling between rooms, a GAT mechanism to automatically learn the strength of thermal coupling, and a temporal network (GRU / DilatedCNN) to capture time dependence and multi-scale dynamics; to achieve prediction of key variables such as room temperature, EFH power demand, and PV available energy for the next 15–240 minutes. The ST-GAT output serves as the "future state auxiliary input" for PERTD3, realizing "predictive-aware RL".

[0083] Strategy Optimization: Energy Regulation Decision Model Based on PERTD3

[0084] In unsteady-state environments, control algorithms capable of handling continuous action spaces, system lag, random disturbances, energy coupling, SOC boundary constraints, and user preference interference are essential. This invention proposes PERTD3 (Prioritized Experience Replay Twin Delayed DDPG), which enhances traditional TD3 as follows: Prioritized Experience Replay (PER) prioritizes high TD-Error samples for learning, improving convergence speed and policy stability. A dual-critic network addresses Q-value overestimation, making it suitable for sub-plateau meteorological conditions with strong random disturbances. Policy Delay prevents frequent policy updates from causing oscillations. Target Smoothing enhances robustness and generalization ability. ST-GAT predictions are introduced as extended states to address the "control blind zone" caused by EFH lag. The actions output by PERTD3 include: electric underfloor heating power; energy storage charging and discharging power; grid power purchase / feedback decisions; achieving multi-objective optimization of energy saving, comfort, cost, and equipment lifespan.

[0085] User comfort is not solely determined by room temperature; it is also influenced by factors such as the rate of temperature change, room usage frequency, time of day preferences, and individual differences. This invention constructs a multi-dimensional user feedback mechanism combining a temperature utility function, subjective ratings, and behavioral recognition. By incorporating this mechanism into a pull-algorithm (RL) reward function, the system can dynamically learn user preferences, achieve personalized energy management, avoid discomfort caused by excessive energy saving, and optimize the user experience in the long term. This results in a truly user-centric energy management system.

[0086] In multi-energy complementary electric underfloor heating systems for rural residences in sub-plateau regions, numerous factors influence indoor temperature changes, photovoltaic power generation fluctuations, and energy storage behavior. These include multi-dimensional variables such as outdoor temperature, wind speed, irradiance, PV module temperature, ground temperature, EFH power, and SOC state, with complex nonlinear coupling relationships among these variables. To identify the key features that most significantly affect system operation and to construct a high-quality input feature space for subsequent models (ST-GAT prediction model and PERTD3 reinforcement learning model), this invention employs Random Forest (RF) to perform feature importance analysis and input variable selection on the measured data.

[0087] Random Forest (RF) is a nonlinear model based on ensemble learning, consisting of multiple regression trees. By introducing mechanisms such as "random sub-feature selection" and "oversampling of different samples," RF effectively avoids overfitting and is suitable for the noisy and highly nonlinear experimental data used in this invention. Its basic idea is to achieve high-precision modeling capabilities through a combination of bootstrap random sampling and multi-tree voting with averaging. The form of the random forest regression model is as follows:

[0088]

[0089] Where: N: the number of trees in the forest; The predicted output X of the i-th regression tree; input feature vector; The final prediction result of the model.

[0090] In a regression tree, each node split changes the mean squared error (MSE) of the samples. RF uses the sum of the MSE reductions caused by a feature across all trees to measure its importance. The importance of feature k is calculated as follows:

[0091]

[0092] in: Importance score of feature k; The set of all nodes that are split using feature k; : Number of node samples; Total number of samples; , The feature that significantly reduces the change in MSE before and after splitting will receive a higher importance score.

[0093] Based on 90 days of actual measurement data (approximately 100,000 minute-level samples), this invention performs feature screening according to the following steps:

[0094] Input variable set construction:

[0095] Based on the collected data from the multi-energy complementary system, a preliminary set of variables was constructed: meteorological characteristics. Energy characteristics Building thermal environment characteristics User characteristics .

[0096] Random Forest Training: A Random Forest regressor consisting of 500 trees is used, with 70% of the data used for model training and 30% for feature evaluation.

[0097] Feature importance ranking: RF outputs the contribution of dominant variables, used to explain system dynamics.

[0098] Feature selection and dimensionality reduction: Based on the importance threshold (5% in this embodiment), variables with little impact on building heat load are removed, and an optimized set of input variables is constructed for: ST-GAT prediction model; PERTD3 reinforcement learning state space.

[0099] Feature Importance Results and Physical Meaning Analysis: The importance ranking of variables obtained from the RF model is shown in the table below. The RF analysis results reveal important laws of system dynamics: external temperature and wind speed are the main driving forces of heat load, which is a typical characteristic of the sub-plateau strong convection environment. Irradiance and backplane temperature jointly determine PV fluctuations, showing why PV temperature-irradiance decoupling is necessary. Ground temperature and EFH power jointly determine the indoor heating rate, confirming that EFH is a high-hysteresis system, making prediction indispensable. SOC is particularly critical for maintaining nighttime temperature, providing a physical basis for RL optimization at night.

[0100]

[0101] Due to the typical winter characteristics of sub-plateau regions—strong radiation, low temperatures, and rapid temperature rise—the output power of photovoltaic (PV) modules exhibits a significant temperature-irradiance coupling effect: increased irradiance leads to increased PV output, but simultaneously, the rising temperature of the module backsheet reduces PV efficiency, causing PV power changes to be out of sync with irradiance. Therefore, directly using… Irradiance, as a predictive and control input, can introduce biases and weaken model stability. Therefore, this invention constructs a PV temperature-irradiance decoupled model based on the aforementioned measured data, decoupling the positive effect of irradiance on PV power from the negative effect of temperature rise on efficiency, thereby obtaining more stable and physically consistent PV input characteristics.

[0102] The physical model basis for PV output:

[0103] The output power of a photovoltaic module can be expressed as:

[0104]

[0105] in: : Irradiance of inclined surface PV area; Component conversion efficiency under the influence of temperature.

[0106] It approximately follows a linear decay relationship with temperature:

[0107]

[0108] in: =25℃; The coefficient is the temperature coefficient (approximately 0.004–0.005 / °C for the PV module of this invention). It can be seen that increased irradiation improves output, but an increase in component temperature (almost simultaneously) offsets some of the power gain.

[0109] Temperature-irradiation coupling characteristics analysis: Regression analysis of measured data reveals that the module temperature rises with a lag of 10–20 minutes after irradiation increases, and the peak irradiation (800 W / m²) is reached later. 2 )hour, It can instantly rise to 35–45°C. For every 10°C increase, PV output decreases by approximately 5–7%. The thin air at high altitudes reduces the backplate's heat dissipation capacity, making thermal coupling more pronounced. This illustrates that PV power is not simply... The linear function of temperature rise will suppress the power generation increase brought about by increased irradiance, and must be implemented. and The decoupling point. Furthermore, this invention constructs a PV decoupling model that conforms to physical mechanisms.

[0110] Construction of the PV temperature-irradiation decoupling factor:

[0111] Based on the PV efficiency model, the effect of module temperature is used as a degradation factor to define temperature-corrected irradiance:

[0112]

[0113] in: ,but = ;like ,but This indicator can effectively compensate for the decrease in efficiency caused by temperature rise, making the irradiation input more in line with physical laws.

[0114] Linear regression model for V power output:

[0115] After constructing the temperature-corrected irradiance, the PV output and the corrected irradiance show an approximately linear relationship:

[0116]

[0117] in: , The coefficients are obtained through least squares fitting. It approaches the combined efficiency of an inverter and its modules. The bias term representing low irradiance (typically close to 0) is obtained through regression fitting of measured data: R 2 The improvement was from 0.87 to 0.94, and the RMSE decreased by 15%–25%, with a significant improvement in predictive stability under high-irradiation mutation conditions.

[0118] Model training process for PV decoupling model:

[0119] Step 1: Data preprocessing, removing invalid points with zero power at night; correction Sensor noise; Perform wavelet smoothing for denoising; unify timestamps.

[0120] Step 2: Temperature-corrected irradiance calculation

[0121]

[0122] Step 3: Establish linear or polynomial regression

[0123] This invention employs a linear model and uses cross-validation to ensure generalization ability.

[0124] Step 4: Model Validation

[0125] Evaluation metrics include: RMSE, MAE, MAPE, R 2 .

[0126] The PV decoupling model serves as a "physical preprocessing layer" within the entire AI control system, playing a crucial role in subsequent steps. It provides more stable input features for ST-GAT, as sudden irradiation fluctuations cause drastic fluctuations in the PV time series, making direct input into ST-GAT prone to error accumulation. The decoupled model... It can significantly reduce noise and improve the discriminative power of prediction models. It avoids PERTD3's misjudgment of available energy. PERTD3 requires accurate estimation of future usable PV energy during decision-making; if the input is excessively influenced by noise, it can lead to: overcharging, over-discharging, and misjudging valley / peak periods; ineffective heating or premature temperature rise failure. The PV decoupling model ensures that RL decisions have a more realistic, smoother, and physically consistent estimate of available energy. It also improves the interpretability of multi-energy flow coupling models; the decoupled PV characteristics have clear physical meaning. The decision was made to use "light input". Determine "efficiency decay" Determining "practically usable irradiation" improves the transparency and interpretability of the overall model.

[0127] Heating systems for rural dwellings in sub-plateau regions are characterized by multi-spatial coupling (multiple rooms), time lag (thermal inertia of underfloor heating), and highly disturbed environments (drastic changes in radiation, wind speed, and external temperature). Traditional prediction methods such as LSTM and CNN struggle to simultaneously characterize the thermal coupling structure and temporal dynamics between multiple areas of a building. To address this issue, this invention constructs the ST-GAT model. It learns the spatial thermal coupling relationships between rooms through graph convolution (GCN / GAT) and learns the temporal dynamics of the heating system through a temporal network (GRU / Dilated CNN), enabling multi-step predictions of room temperature, load, EFH power demand, and available PV energy for the next 15–240 minutes.

[0128] The ST-GAT model structure consists of three core modules: a spatial graph convolutional layer, used to capture the thermal coupling relationships between rooms (input node feature matrix, output spatially aggregated node representation); a temporal convolutional / recurrent layer, used to model the underfloor heating hysteresis effect and multi-step dynamic changes (input spatial feature matrix, output temporally encoded node features); and a prediction layer, used to output future multi-step prediction sequences. Outputs include future room temperature sequences, EFH power sequences, and PV available energy sequences (for RL).

[0129] Each room in the farmhouse is considered a node. Construct a space thermal coupling diagram:

[0130]

[0131] in: :N room nodes; The set of edges between nodes; is the adjacency matrix, representing the thermal coupling strength.

[0132] (1) Method for constructing adjacency matrix A

[0133] The adjacency matrix reflects the strongest thermal coupling between the living room and bedroom. South-facing rooms are significantly affected by radiation and contribute greatly to the overall temperature. North-facing rooms dissipate more heat, resulting in a moderate coupling weight. Based on the measured relationship between room temperature and floor temperature, the coupling weight between rooms is defined as follows:

[0134]

[0135] in: Structural distances between rooms (calculated based on the BIM model); Attenuation coefficient (based on actual measurement fitting).

[0136] The core of the GAT layer is to enable the model to automatically learn the thermal coupling strength between rooms. No predefined thermal coupling weights are required; the model can learn on its own. It can identify "strongly coupled rooms" and "weakly coupled rooms" and adapt to coupling changes caused by sudden drops in outside temperature during winter.

[0137] (1) Linear transformation

[0138] Features of each node After linear transformation:

[0139]

[0140] (2) Attention score

[0141] Calculate the attention weights between node i and its neighbor node j:

[0142] )

[0143] in: Learnable attention vectors; Feature splicing; : The thermal coupling weight of node j to node i (unnormalized).

[0144] Normalization

[0145]

[0146] The updated room node characteristics are:

[0147]

[0148] The time modeling module is necessary because underfloor heating has a significant thermal inertia lag of 30–50 minutes, requiring a time model to capture dynamic changes. This invention employs a hybrid structure of Dilated Causal Convolution + GRU, combining long memory capacity with computational efficiency.

[0149] Dilated convolution is used to expand the receptive field, making it suitable for capturing multi-scale behaviors such as slow changes in room temperature and rapid fluctuations in PV output, and for expressing long-term dependencies.

[0150]

[0151] in: : Coefficient of thermal expansion; : Kernel size.

[0152] The GRU can handle temperature hysteresis, the delay relationship between EFH power and room temperature, the indirect impact of PV output abrupt changes on room temperature dynamics, and further capture time state:

[0153]

[0154]

[0155]

[0156]

[0157] The inputs and outputs of ST-GAT are constructed based on the real data measured above. The input features of ST-GAT have been filtered for importance by RF.

[0158] Input features (X)

[0159] Input vector for each room node:

[0160]

[0161] in: Effective irradiation after processing by the PV decoupling model.

[0162] Output feature (Y)

[0163] The prediction results are mainly used to determine in advance whether heating is needed (lag compensation), whether pre-charging during off-peak hours is needed, and to determine the potential output of PV, providing a "predictive state" for RL to PERTD3, and predicting the following for the next H time steps (H=15–240min):

[0164]

[0165] The loss function and training strategy employ weighted multi-task loss:

[0166]

[0167] Each part uses a hybrid loss of MAE+MSE:

[0168]

[0169] After completing feature selection (RF), PV temperature-irradiance decoupling, and ST-GAT multi-step prediction, it is necessary to construct an intelligent control strategy capable of real-time decision-making in complex and unsteady-state environments. Based on the TD3 algorithm, this invention introduces priority experience playback, safety constraints, and prediction information to propose the PERTD3 control model for optimizing the operation strategy of multi-energy complementary electric underfloor heating systems in sub-plateau rural residences. PERTD3 formalizes the heating control problem as a decision-making process in a continuous state-action space, using "energy cost + thermal comfort + energy storage life + peak-valley electricity price response + user feedback" as comprehensive optimization objectives to achieve a balance between energy saving, comfort, and economy.

[0170] The MDP formalization of the heating regulation problem discretizes the system into a time series with a step size of 1 minute, defining the state. : Perceived information of the system at time t (environment + energy + prediction); Action The control variable adjusted at time t, and the state transition: Instant rewards: .

[0171] The goal of reinforcement learning is to maximize the discounted reward:

[0172]

[0173] in As a strategy, This is the discount factor.

[0174] Combining the above measured data with the RF feature importance analysis in 3.2 and the ST-GAT prediction results in 3.4, the state vector at time t is defined: by embedding the ST-GAT prediction results into the state space, "prediction-control coupling" is achieved, which helps to alleviate the problems caused by EFH thermal inertia and system response lag.

[0175]

[0176] in: : Indoor temperature of the i-th room; Ground temperature; Outdoor temperature and wind speed; Current PV output and effective irradiance after decoupling; Energy storage state of charge; Current power of electric underfloor heating; : Future HHH step room temperature / load forecasts provided by ST-GAT (key dimensions selected and pieced together); User-defined target temperature; User comfort feedback (or statistics of historical ratings).

[0177] PERTD3 outputs the continuous control actions to be performed at time t:

[0178]

[0179] in: : Power setting for electric underfloor heating (can be considered as duty cycle or power setting value) Energy storage charging power setting; Energy storage discharge power setting.

[0180] After the action is executed, the system evolves according to the following mechanism: EFH power change → affects future room temperature and ground temperature through the building heat balance equation; ESS charging and discharging power change → determines SOC evolution through energy balance; PV output + electricity price → determines the amount of electricity purchased and the cost. When constraints are violated (such as SOC exceeding limits), stronger penalties are imposed in the environment, while simultaneously trimming to the safety boundary. SOC updates follow:

[0181]

[0182] in: For charge / discharge efficiency; Rated capacity; The time interval is 1 minute.

[0183] To comprehensively reflect comfort, energy consumption, economy, energy storage life, and user feedback, this invention constructs a polynomial reward function:

[0184]

[0185] in These are weighting coefficients, which can be optimized through experiments.

[0186] Let the comfort zone be... The penalty is based on temperature deviation; the greater the deviation of the room temperature from the set temperature, the greater the penalty.

[0187]

[0188] Considering the power consumption and electricity purchase cost of EFH:

[0189]

[0190] in To purchase electricity from the grid, It can be time-of-use electricity pricing.

[0191] Suppressing deep charge / discharge and high C-value operation:

[0192]

[0193] Encourage charging during off-peak hours and minimize electricity purchases during peak hours:

[0194]

[0195] To improve sample utilization efficiency and convergence speed, this invention introduces priority experience replay and uses the TD error magnitude as the sampling weight.

[0196] Sampling probability

[0197]

[0198] in: The TD error of the i-th empirical rule; : Controls the priority level (0 indicates uniform sampling); : Prevent smoothing terms with a probability of 0.

[0199] To reduce the impact of biased sampling on the estimation, weights are introduced:

[0200]

[0201] in: Used for weighted TD error; With training gradually from Increase to 1.

[0202] The overall training steps of PERTD3 can be summarized as follows: Initialize the parameters of the Actor, Critic, and corresponding target networks; initialize the priority experience replay buffer; use the current policy to interact with the environment during interactive sampling; and collect (…). (st,at,rt,st+1)(s_t,a_t,r_t,s_{t+1})(st,at,rt,st+1) and store in the PER buffer. Sampling batch by priority: Sample from the buffer according to P(i) and calculate the corresponding IS weights wi. Calculate the TD target and update the Critic: Calculate yty_tyt and update the two Critics with weighted loss. Update the Actor and target network every d steps: Update the Actor according to the TD3 method and softly update the target network. Update priority: Update the empirical priority according to the new TD error. Repeat steps 2–6 until convergence or the set number of rounds is reached.

[0203] Although indoor temperature is a core indicator for measuring heating comfort, the subjective experience of actual users is influenced by a variety of factors, including the rate of temperature change, historical comfort experience, usage habits, and time preferences (such as different temperature sensitivities in the morning and evening). Traditional control methods based on fixed set temperatures (such as 20 °C) or PMV models are difficult to reflect these user differences, and may easily lead to situations where "energy is saved but the experience is poor" or "the experience is good but energy is wasted."

[0204] To achieve integrated control of "people-equipment-energy", this invention constructs a user feedback mechanism model, including: a temperature utility function, a user preference learning module, and a user feedback integrated into the PERTD3 reward function (RL-Feedback Coupling). This mechanism enables the control system to continuously adjust the heating strategy based on the user's actual experience, achieving true long-term comfort optimization and personalized heating.

[0205] This invention introduces a Gaussian temperature utility function:

[0206]

[0207] Where: TTT: Current indoor temperature; Tpref: User-preferred temperature (dynamically changing); σ: Temperature tolerance.

[0208] Model physical meaning: when hour:

[0209]

[0210] Users reach their optimal comfort state. As the temperature deviates further from the preferred temperature, utility gradually decreases, with the rate of decrease determined by σ. Highly sensitive users (such as the elderly and children) have smaller σ; while younger users have higher tolerance (and larger σ). This reflects the human perception pattern of "the greater the deviation, the greater the penalty."

[0211] User comfort experience is not constant throughout the heating season, but rather changes dynamically with behavior, habits, and even external weather. Therefore, this invention designs an online user preference learning model to adjust Tpref and σ in real time.

[0212] Users can provide real-time comfort ratings via mobile app or smart control panel. .

[0213] Preferred temperatures will be updated based on ratings:

[0214]

[0215] Where: ηT\eta_TηT: learning rate; when the user rating is higher than the system's predicted utility → preference moves closer to the current temperature; when the rating is lower → preference moves away from the current temperature.

[0216] Temperature tolerance (σ) update mechanism

[0217] If the rating has fluctuated significantly recently, it indicates that the user is sensitive to temperature changes, and σ needs to be reduced.

[0218]

[0219] in: The past N ratings; The magnitude of fluctuation reflects stability.

[0220] More stable user behavior → Increased σ\sigmaσ

[0221] User sensitivity → σ\sigmaσ reduces default behavior inference (no need for active user feedback).

[0222] When users do not actively provide ratings, behavioral data is used to infer: going out → reduced room temperature requirement; falling asleep → reduced preferred temperature (e.g., a lower nighttime temperature pattern); waking up → increased preferred temperature (increased thermal comfort requirement); this mechanism is implemented through a weighted average of the latest behavioral window.

[0223]

[0224] User feedback is coupled with the PERTD3 reward function (RL–Feedback Coupling)

[0225] In order for reinforcement learning strategies to proactively optimize in the direction that "users prefer", user utility needs to be incorporated into the reward function of PERTD3.

[0226] The PERTD3 reward function is defined in Section 3.5 as follows:

[0227]

[0228] User feedback items:

[0229]

[0230] The user utility of the current temperature Long-term average utility (benchmark): If the user experience is better than expected, reward it; if the experience is worse than the long-term level, punish it.

[0231] The basic principles and main features of the present invention have been described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are only illustrative of the principles of the present invention. Various changes and modifications can be made to the present invention without departing from the spirit and scope of the present invention. All such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the invention is defined by the appended claims and their equivalents.

Claims

1. A multi-energy complementary electric underfloor heating intelligent predictive control system, characterized in that, include: The module includes a perception module, a prediction module, an optimization and control module, and a user feedback module. The sensing module is used to collect meteorological data, photovoltaic power generation data, energy storage system data, electric floor heating system data, indoor temperature data, and user settings and behavior data to construct a multi-source high-resolution dataset. The prediction module includes a feature filtering unit, a photovoltaic output prediction unit, and a load multi-scale prediction unit. The feature filtering unit identifies the dominant features affecting system operation based on a random forest model. The photovoltaic output prediction unit predicts photovoltaic output based on a decoupled model of photovoltaic module temperature and irradiance. The load multi-scale prediction unit is based on a spatiotemporal graph attention network model, which integrates building spatial topology and time series to achieve multi-step prediction of indoor temperature and load. The optimization and control module is based on an improved reinforcement learning algorithm. It takes the output of the prediction module as input, constructs a multi-objective reward function that includes energy consumption, comfort, system stability and economy, and outputs continuous control commands for electric underfloor heating power, energy storage system charging and discharging power and grid interaction power. The user feedback module is used to collect user comfort evaluation and behavior data, build a user preference model, and dynamically integrate user feedback information into the reward function of the optimization and control module to achieve adaptive adjustment of the heating strategy.

2. The intelligent predictive control system for multi-energy complementary electric underfloor heating according to claim 1, characterized in that, The sensing module includes an Internet of Things (IoT) sensor network deployed inside and outside the building and on the side of energy equipment. The sensor network collects data at a resolution of no less than 1 minute and performs data cleaning, alignment and storage through edge computing and cloud platform.

3. The intelligent predictive control system for multi-energy complementary electric underfloor heating according to claim 1, characterized in that, The decoupling model used in the photovoltaic power output prediction unit is as follows: ； in: ,but = ;like ,but This indicator can effectively compensate for the decrease in efficiency caused by temperature rise, making the irradiation input more in line with physical laws.

4. The intelligent predictive control system for multi-energy complementary electric underfloor heating according to claim 1, characterized in that, The load multi-scale prediction unit uses a spatiotemporal graph attention network model to construct a spatial graph with each room in the building as a node. It dynamically learns the thermal coupling weights between rooms through a graph attention mechanism and uses a temporal network combining dilated causal convolution and gated recurrent units to capture the multi-scale temporal dynamics of the system. It outputs the predicted sequence of indoor temperature, electric underfloor heating power demand and photovoltaic available energy for each room in the next 15 to 240 minutes.

5. The intelligent predictive control system for multi-energy complementary electric underfloor heating according to claim 1, characterized in that, The improved reinforcement learning algorithm used in the optimization and control module is a dual-delay deep deterministic strategy gradient algorithm based on priority experience replay. Its state space includes the current indoor temperature, outdoor temperature, wind speed, effective irradiance, energy storage state of charge, electric underfloor heating power, user-set temperature, user feedback information, and future multi-step prediction sequence output by the load multi-scale prediction unit. Its action space consists of continuous value control commands for electric underfloor heating set power, energy storage charging power, and energy storage discharging power.

6. The intelligent predictive control system for multi-energy complementary electric underfloor heating according to claim 5, characterized in that, The multi-objective reward function of the optimization and control module is composed of the following weighted summation: ； in These are weighting coefficients, which can be optimized through experiments.

7. The system according to claim 1, characterized in that, The user preference model of the user feedback module is constructed based on a Gaussian temperature utility function. ； Where: T: current indoor temperature; Tpref: user preferred temperature (dynamically changing); σ: temperature tolerance.

8. A method for intelligent predictive control of multi-energy complementary electric underfloor heating, characterized in that, The multi-energy complementary electric underfloor heating intelligent predictive control system applied to any one of claims 1-7 includes the following steps: S1: Continuously collect multi-source operational data through the perception module, and perform data cleaning and preprocessing to build a high-quality training and inference dataset; S2: Use the random forest model to perform feature importance analysis on the dataset and select the dominant features; S3: Based on the decoupled model of photovoltaic module temperature and irradiance, predict photovoltaic output; S4: Based on the spatiotemporal graph attention network model, it integrates building space topology and historical time series data to achieve prediction of indoor temperature and system load in multiple future steps; S5: Based on user comfort evaluation and behavioral data, dynamically update the preference temperature and tolerance parameters in the user preference model; S6: Using the features selected in step S2, the prediction results of steps S3 and S4, and the user preference information in step S5 as state inputs, the reinforcement learning optimization control model based on priority experience replay is used to calculate and output the optimal electric underfloor heating power, energy storage charging and discharging power, and grid interaction power control commands. S7: Executes control commands and continuously updates the prediction model and optimizes the control strategy based on system operation results and user feedback.

Images