Renewable energy heating group control method and system based on machine learning

Abstract

The invention relates to the technical field of renewable energy heating, solves the problem that the existing renewable energy heating technology lacks comprehensive modeling and collaborative optimization of multi-source environmental data, user demand characteristics and equipment group operation states, and provides a renewable energy heating group control method and system based on machine learning. The method comprises the steps that multi-source environment data, heating demand information and operation state parameters of a to-be-heated area are obtained; processing the multi-source environment data, the heating demand information and the operation state parameters by using a first machine learning model to obtain a target heat load prediction result and a productivity prediction result; determining a group control strategy of the renewable energy equipment group according to a second machine learning model in combination with the target thermal load prediction result and the productivity prediction result; and according to the group control strategy, the renewable energy equipment group is controlled, and heating of the to-be-heated area is completed. The cooperative efficiency and the operation effect of renewable energy source heating are improved.

Description

Technical Field

[0001] This invention relates to the field of renewable energy heating technology, and in particular to a method and system for group control of renewable energy heating based on machine learning. Background Technology

[0002] In recent years, with the continuous advancement of policies promoting energy conservation, emission reduction, and green development, heating methods based on renewable energy have gradually become an important alternative to traditional fossil fuel heating. Clean energy sources such as solar, wind, biomass, and geothermal energy are increasingly being used in heating systems. These heating methods not only reduce carbon dioxide emissions but also improve energy efficiency, thereby meeting the needs of clean heating and low-carbon development.

[0003] Existing renewable energy heating technologies mainly include solar thermal heating, air-source or ground-source heat pump heating, and wind power and biomass boiler-assisted heating. These technologies typically collect environmental parameters such as temperature and humidity through sensors, and then combine this with power adjustment of equipment to meet the heating needs of local buildings or areas. However, current renewable energy heating technologies mostly rely on fixed regulation logic or simple predictive models, often scheduling based on a single energy type. They lack comprehensive modeling of multi-source environmental data, user demand characteristics, and the operating status of equipment groups, making it difficult to fully realize the potential of synergistic heating from multiple renewable energy sources.

[0004] Therefore, how to achieve coordinated optimization of renewable energy heating based on comprehensive modeling of multi-source environmental data, user demand characteristics, and equipment group operating status is an urgent problem to be solved. Summary of the Invention

[0005] In view of this, the present invention provides a machine learning-based method and system for the group control of renewable energy heating, which solves the problem of existing renewable energy heating technologies lacking multi-source environmental data, comprehensive modeling and collaborative optimization of user demand characteristics and equipment group operating status.

[0006] The technical solution adopted in this invention is: In a first aspect, the present invention provides a method for group control of renewable energy heating based on machine learning, the method comprising: Acquire multi-source environmental data of the area to be heated, user heating demand information, and operating status parameters of renewable energy equipment groups used for heating; Using a pre-trained first machine learning model, the multi-source environmental data, the heating demand information, and the operating status parameters are processed to obtain the target heat load prediction results for the area to be heated and the capacity prediction results for the renewable energy equipment group within a preset time period. Based on the preset second machine learning model, and combining the target heat load prediction results and the capacity prediction results, the group control strategy for the renewable energy equipment group is determined; According to the group control strategy, the renewable energy equipment group is controlled to complete the heating of the area to be heated.

[0007] Preferably, the acquisition of multi-source environmental data of the area to be heated, user heating demand information, and operating status parameters of the renewable energy equipment group used for heating includes: By deploying environmental sensors inside and outside the area to be heated, the first environmental parameters of multiple preset local areas within the area to be heated and the second environmental parameters within a preset area outside the area to be heated are collected. Based on the preset weights corresponding to each preset local region, the first environmental parameters are weighted and fused to obtain the multi-source environmental parameters within the area to be heated. The second environmental parameter and the multi-source environmental parameter are subjected to time-series synchronization processing to obtain the multi-source environmental data; The start-stop status, output power, and operating efficiency data of the renewable energy equipment group are collected to obtain the operating status parameters; Based on the target set temperature and desired heating time input by the user, and combined with the user's historical heating preference information, the heating demand information is obtained.

[0008] Preferably, the step of using a pre-trained first machine learning model to process the multi-source environmental data, the heating demand information, and the operating status parameters to obtain the target heat load prediction results for the area to be heated and the capacity prediction results for the renewable energy equipment group within a preset time period includes: The multi-source environmental data, the heating demand information, and the operating status parameters are preprocessed to obtain preprocessed target environmental data, target heating demand information, and target operating status parameters. Based on the target environmental data, target heating demand information, and target operating status parameters, extract the first feature information for heat load prediction and the second feature information for production capacity prediction. The first feature information is subjected to time-series expansion, lag term construction and normalization to obtain the first model input feature adapted to heat load prediction; The second feature information is smoothed, trend extracted, and seasonal factor labeled to obtain the second model input features adapted to capacity forecasting. The first model input features are input into the first machine learning model for inference processing to obtain the target heat load prediction result; The second model's input features are input into the pre-trained first machine learning model for inference processing to obtain the production capacity prediction result.

[0009] Preferably, the step of extracting first feature information for heat load prediction and second feature information for production capacity prediction based on the target environmental data, target heating demand information, and target operating status parameters includes: Based on the target environmental data, time series sampling and moving average processing are performed on the environmental temperature, environmental humidity, wind speed and solar radiation intensity to obtain a subset of environmental features characterizing external climate conditions. Based on the target heating demand information, the target set temperature, expected heating time and historical heating records are marked by time period and heating habits are summarized to obtain a subset of demand features that characterize user demand. Based on the target operating state parameters, the equipment start-up and shutdown status, output power curve and operating efficiency index are screened and normalized to obtain a subset of operating features that characterize the equipment's operating characteristics. The environmental feature subset, demand feature subset and operation feature subset are correlated and analyzed to extract the first feature information, which includes the temperature sequence outside the area to be heated, the target temperature sequence inside the area to be heated and the historical load difference sequence. The environmental feature subset and the operational feature subset are correlated and analyzed to extract the second feature information, which includes solar radiation intensity sequence, wind speed sequence and equipment operation efficiency index sequence.

[0010] Preferably, the pre-training process of the first machine learning model includes: A training sample set is constructed based on pre-collected historical multi-source environmental data, historical heating demand information, and corresponding historical operating status parameters of renewable energy equipment; The historical multi-source environmental data is denoised, missing values ​​are imputed, and timestamps are aligned to obtain the environmental time series for training. The historical heating demand information is formatted, normalized, and labeled with user habit features to obtain the demand time series for training. The historical operating state parameters are smoothed, standardized, and efficiency factors are calculated to obtain the equipment operating time series for training. Based on the environmental time series, demand time series, and equipment operation time series, first training feature information for heat load prediction and second training feature information for capacity prediction are extracted respectively. The first training feature information is matched with the actual historical heat load observations to construct the first training data pair; The second training feature information is matched with the actual observed values ​​of equipment capacity to construct a second training data pair; The first training data pair and the second training data pair are respectively input into the first machine learning model for supervised training until the prediction error of the model meets the preset convergence condition, thus obtaining the first machine learning model after training.

[0011] Preferably, determining the group control strategy for the renewable energy equipment group based on a preset second machine learning model, combined with the target heat load prediction result and the capacity prediction result, includes: Based on the target heat load prediction results, obtain the heat demand information of the area to be heated in each prediction period; Based on the capacity forecast results, obtain the energy availability information of the renewable energy equipment group in each forecast period; Obtain a pre-established group control optimization objective function, wherein the objective function includes at least a heating comfort objective, an energy utilization objective, and an operating cost objective; Based on the heat demand information and the available energy information, construct supply and demand balance constraints; Based on the preset start-stop logic, output power limit, and operating efficiency boundary of the renewable energy equipment group, construct the equipment operation constraints; The supply and demand balance constraints and equipment operation constraints are input into the second machine learning model, and combined with the group control optimization objective function, multi-objective optimization is performed to obtain the solution result; Based on the solution results, the start / stop commands and power allocation schemes for the device group in each prediction period are obtained, thus obtaining the group control strategy.

[0012] Preferably, the step of inputting the supply-demand balance constraints and equipment operation constraints into the second machine learning model, and combining them with the group control optimization objective function to perform multi-objective optimization and obtain the solution results includes: Using the supply-demand balance constraints and the equipment operation constraints as input constraints, and combining them with the group control optimization objective function, a group control optimization problem with multiple constraints is constructed. The second machine learning model is used to iteratively solve the group control optimization problem. In each iteration, the start / stop commands and power allocation parameters of the device group are adjusted to obtain multiple candidate group control strategies. The heating comfort, energy efficiency and operating cost of each candidate group control strategy are comprehensively scored, and the candidate group control strategies are ranked according to the scoring results to obtain the ranking results; Based on the ranking results, the optimal candidate group control strategy is selected as the solution result.

[0013] Preferably, the pre-training process of the second machine learning model includes: Based on the pre-collected historical target heat load prediction results and historical production capacity prediction results, combined with historical heating comfort index, energy utilization rate index and operating cost data, a multi-objective training sample set is constructed. The historical heat load prediction results and historical production capacity prediction results are normalized and time-series aligned to obtain the input feature sequence for training. The historical heating comfort index, energy utilization rate index and operating cost data are standardized and weighted to obtain a multi-objective output sequence for training. By mapping the input feature sequence to the multi-target output sequence, training data pairs are constructed. The training data is input into the second machine learning model, and a multi-objective optimization algorithm is used for supervised training. During the training process, the heating comfort target, energy utilization target and operating cost target are balanced through dynamic weight allocation. The training continues iteratively until the combined error of the second machine learning model on the validation set meets the preset convergence condition, thus obtaining the trained second machine learning model.

[0014] In a second aspect, the present invention provides a renewable energy heating group control system based on machine learning, comprising: at least one processor, at least one memory, and computer program instructions stored in the memory, wherein when the computer program instructions are executed by the processor, the method of the first aspect described above is implemented.

[0015] In summary, the beneficial effects of the present invention are as follows: This invention provides a machine learning-based renewable energy heating group control method and system. The method includes: acquiring multi-source environmental data of the area to be heated, user heating demand information, and operating status parameters of a group of renewable energy equipment used for heating; processing the multi-source environmental data, heating demand information, and operating status parameters using a pre-trained first machine learning model to obtain a target heat load prediction result for the area to be heated and a capacity prediction result for the renewable energy equipment group within a preset time period; determining a group control strategy for the renewable energy equipment group based on a preset second machine learning model, combined with the target heat load prediction result and the capacity prediction result; and controlling the renewable energy equipment group according to the group control strategy to complete the heating of the area to be heated. This invention first collects and fuses multi-source environmental data such as indoor and outdoor temperature, humidity, wind speed, and solar radiation intensity through sensors, and simultaneously combines user-set temperature, desired heating time, and historical heating preference information, as well as the start / stop status, output power, and operating efficiency parameters of the equipment group, to construct comprehensive input features. Subsequently, the first machine learning model was used to accurately predict heat load and production capacity, providing forward-looking demand and supply references for the heating system. Furthermore, the prediction results were combined with supply and demand balance constraints and equipment operation constraints, and a second machine learning model was used to perform multi-objective optimization. Taking into account heating comfort, energy utilization rate and operating cost, the optimal group control strategy was generated. This achieved unified modeling and intelligent optimization scheduling of multi-source data, user characteristics and equipment status, thereby significantly improving the collaborative efficiency and operational effect of renewable energy heating. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, and these are all within the protection scope of the present invention.

[0017] Figure 1 This is a schematic diagram illustrating the overall workflow of the machine learning-based renewable energy heating group control method in Embodiment 1 of the present invention. Figure 2 This is a flowchart illustrating the process of determining the group control strategy for the renewable energy equipment group in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the structure of the renewable energy heating group control system based on machine learning in Embodiment 2 of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. In the description of the present invention, it should be understood that the terms "center," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, the element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. Where there is no conflict, embodiments of the present invention and the various features thereof can be combined with each other, all of which are within the scope of protection of the present invention.

[0019] Example 1 Please see Figure 1 Embodiment 1 of this invention discloses a machine learning-based renewable energy heating group control method, the method comprising: Acquire multi-source environmental data of the area to be heated, user heating demand information, and operating status parameters of renewable energy equipment groups used for heating; Specifically, multi-source environmental data refers to exemplary sensor data, including temperature, relative humidity, wind speed, solar irradiance, and air quality inside and outside the area to be heated; heating demand information includes user-set temperature and desired heating time; and operating status parameters include equipment start / stop status, instantaneous output power, operating efficiency, and energy storage charge / heat charge status. The purpose of this step is to establish a complete and time-consistent input foundation to support subsequent prediction and optimization. This is achieved by deploying sensors and gateways inside and outside the area, sampling, denoising, filling in missing values, and aligning timestamps according to a unified time scale, and performing partitioned weighted aggregation, data caching, and quality verification at the edge or cloud. Simultaneously, demand and operating parameters are periodically retrieved and formatted from user terminals and equipment controllers and written to a historical database for batch reading, thereby ensuring the integrity and availability of input data, reducing prediction and scheduling uncertainty, facilitating timely anomaly detection, and supporting personalized control.

[0020] Using a pre-trained first machine learning model, the multi-source environmental data, the heating demand information, and the operating status parameters are processed to obtain the target heat load prediction results for the area to be heated and the capacity prediction results for the renewable energy equipment group within a preset time period. Specifically, the first machine learning model is a time-series forecasting model or an ensemble forecaster, which outputs the target heat load forecast results for several future time periods and the available capacity of various renewable energy devices in the same time period as the capacity forecast results. The purpose is to transform historical and real-time information into forward-looking quantitative supply and demand forecasts, providing a basis for scheduling decisions. The process of processing the multi-source environmental data, the heating demand information, and the operating status parameters includes: constructing model input features, taking different input paths for heat load and capacity, performing batch / streaming inference, and generating forecasts with uncertainty; in terms of deployment, cloud training, edge inference, and online correction mechanisms can be adopted to provide accurate and time-resolution demand and supply previews, reduce mismatches caused by short-term fluctuations, and improve the utilization rate of renewable energy.

[0021] Based on the preset second machine learning model, and combining the target heat load prediction results and the capacity prediction results, the group control strategy for the renewable energy equipment group is determined; Specifically, the second machine learning model refers to the decision model used to map predictions to scheduling strategies, including trained optimization agents, policy networks, or learning-based MPC. The group control strategy refers to the start-up, shutdown, output allocation, and energy storage charging / discharging commands for each device within each prediction period. The goal is to generate a safe and executable overall scheduling scheme by balancing and satisfying multiple objectives (comfort, energy utilization, and operating costs) and constraints, while ensuring comfort. In practice, the prediction results must first be transformed into supply-demand differences, supply-demand balance and device capacity constraints must be constructed, and then these constraints and a predefined objective function are used as input to call the second model. The model can generate candidate strategies, perform rapid simulation evaluation, or directly output parameterized strategies. Before output, feasibility checks and rule filtering should be performed, and a timestamped group control command sequence and backup strategies should be generated to achieve coordinated scheduling from a global perspective, reducing energy consumption and costs, improving comfort, and avoiding frequent switching or overload of local devices.

[0022] According to the group control strategy, the renewable energy equipment group is controlled to complete the heating of the area to be heated.

[0023] Specifically, the equipment start / stop instructions and power allocation schemes corresponding to each predicted time period in the group control strategy are transformed into specific control commands and sent to the control terminal of the renewable energy equipment group. During execution, the start / stop status, output power, and operating efficiency of the renewable energy equipment group are monitored in real time to ensure that each device operates according to the group control strategy. When a deviation in equipment execution or a sudden change in the external environment is detected, the dynamic adjustment mechanism of the group control strategy is triggered to readjust the start / stop instructions and power allocation schemes to maintain supply and demand balance. Finally, through the coordinated control of the renewable energy equipment group, the heat demand of the area to be heated is guaranteed to be met, achieving a heating effect that balances energy saving, comfort, and operational safety.

[0024] Preferably, the acquisition of multi-source environmental data of the area to be heated, user heating demand information, and operating status parameters of the renewable energy equipment group used for heating includes: By deploying environmental sensors inside and outside the area to be heated, the first environmental parameters of multiple preset local areas within the area to be heated and the second environmental parameters within a preset area outside the area to be heated are collected. Specifically, environmental sensors include, but are not limited to, temperature sensors, relative humidity sensors, anemometers, light / solar irradiance sensors, and air quality sensors both inside and outside the area to be heated. Their purpose is to obtain raw observations covering the local microclimate and surrounding meteorological conditions, serving as the foundational input for subsequent forecasting and scheduling. In practice, sensor nodes are deployed according to preset local areas (e.g., floors, machine rooms, unit rooms, or pipe network zones). These nodes periodically report first and second environmental parameters, including sensor ID, area ID, and timestamp, via wired or wireless gateways. Simultaneously, basic calibration, redundant deployment, and anomaly alarms (such as sensor disconnection or out-of-range values) are performed at the acquisition end. Sampling frequency, sampling accuracy, and transmission delay are incorporated into metadata for subsequent data quality control, thereby constructing a fine-grained and traceable environmental observation system that provides a reliable raw data source for accurate heat load and capacity assessment.

[0025] Based on the preset weights corresponding to each preset local region, the first environmental parameters are weighted and fused to obtain the multi-source environmental parameters within the area to be heated. Specifically, multiple local observations are merged into a summary that reflects the state of the entire area to be heated, in order to reduce the impact of single-point anomalies and reflect the regional thermal characteristics. According to a predefined weight table, the weights can be set based on the area, population density, thermal characteristics of the building envelope, or sensor reliability. The first environmental parameters of the same type are weighted and averaged according to the time synchronization window. When extreme values ​​are truncated or replaced, if necessary, they are not. When some sensors are missing, the weights are redistributed according to normalized weights or interpolated from neighboring areas. The source of the weights and the reliability index are recorded during the fusion process for auditing and subsequent weight adjustment. Multi-source environmental parameters reflecting the thermal state of the entire area are generated, which takes into account spatial heterogeneity and improves the overall data robustness, which is beneficial to subsequent feature construction and prediction accuracy.

[0026] The second environmental parameter and the multi-source environmental parameter are subjected to time-series synchronization processing to obtain the multi-source environmental data; Specifically, data from internal and external sources, at different frequencies and with varying time delays are unified to a single time base, forming a time-series dataset directly usable for modeling. The process includes resampling and interpolating the second environmental parameter and the multi-source environmental parameters at a unified time scale (e.g., every 15 minutes or hour); correcting clock drift using timestamp correction (based on NTP or gateway clock calibration); buffering delayed data and merging it according to arrival order or priority; simultaneously, filling in missing sampling points using nearest-neighbor interpolation or model prediction and marking the filling points in the data; finally, the synchronized second environmental parameter and multi-source environmental parameters are written into a time-series database or data lake using a time-series index to obtain the multi-source environmental data and generate corresponding data quality reports. This provides a consistent, low-latency, and auditable multi-source environmental data time series, providing a stable input basis for heat load and capacity models and reducing prediction errors caused by time misalignment.

[0027] The start-stop status, output power, and operating efficiency data of the renewable energy equipment group are collected to obtain the operating status parameters; Specifically, this step is used to monitor the real-time operating status of various renewable energy devices (such as solar inverters, wind turbines, heat pumps, and thermal / energy storage devices) as an important basis for capacity assessment and scheduling constraints. It periodically retrieves start / stop status signals, instantaneous output power readings, and operating efficiency calculated or estimated based on sensors / controllers through the communication interface with the field controller or inverter. The collected data is then filtered, averaged, and anomaly flagged (such as sudden changes, disconnections, or exceeding limits). Simultaneously, the available power range and short-term adjustable range are calculated based on the equipment's rated parameters. This information is then compiled into timestamped operating status parameters for optimization purposes, providing accurate equipment-side capability and constraint information, supporting feasibility verification, constraint modeling, and secure and reliable group control deployment.

[0028] Based on the target set temperature and desired heating time input by the user, and combined with the user's historical heating preference information, the heating demand information is obtained.

[0029] Specifically, the system integrates users' immediate needs with long-term usage habits into a demand-side description that can be used for aggregated scheduling. It receives the target set temperature and desired heating time from the user end (such as a smart thermostat, mobile APP, or building management system) and reads the user's or zone's historical heating preference information (including past set temperature distribution, commonly used heating periods, manual overwrite records, and abandonment rules) from the historical database. Then, it merges the immediate input and historical preferences according to a priority strategy (e.g., immediate explicit input is given priority, or historical typical settings are used when there is no immediate input), and generates structured heating demand information containing target temperature, heating start and end time, acceptable temperature deviation, and priority identifier. At the same time, it records whether it is a temporary overwrite for subsequent strategy adjustment. This provides a personalized and executable demand profile for prediction and scheduling, improves user comfort satisfaction rate, and facilitates the consideration of individual differences and overall optimization in group control.

[0030] Preferably, the step of using a pre-trained first machine learning model to process the multi-source environmental data, the heating demand information, and the operating status parameters to obtain the target heat load prediction results for the area to be heated and the capacity prediction results for the renewable energy equipment group within a preset time period includes: The multi-source environmental data, the heating demand information, and the operating status parameters are preprocessed to obtain preprocessed target environmental data, target heating demand information, and target operating status parameters. Specifically, multi-source environmental data, heating demand information, and operational status parameters are first cleaned and standardized: this includes noise reduction, outlier detection and removal, missing value imputation (interpolation or replacement with adjacent time periods), unit unification and dimensional conversion, timestamp alignment and resampling, and generation of quality identifiers (such as credibility and imputation markers). Demand and operational data are also formatted and standardized (e.g., user input is standardized into structured records). In implementation, batch processing and streaming preprocessing can be completed in an edge gateway or cloud pipeline, and the results are written to a time-series database. The output preprocessed target environmental data, target heating demand information, and target operational status parameters are guaranteed to be time-series consistent, dimension-unified, and include data quality metadata to reduce subsequent modeling errors and improve system robustness.

[0031] Based on the target environmental data, target heating demand information, and target operating status parameters, extract the first feature information for heat load prediction and the second feature information for production capacity prediction. Specifically, using pre-processed data, two types of candidate features are constructed: Examples of the first feature information for heat load prediction include: the temperature sequence outside the area to be heated, the target temperature sequence within the area to be heated, historical load baseline, indoor-outdoor temperature difference, equivalent parameters of the building envelope, and the average temperature difference within the time window; examples of the second feature information for capacity prediction include: solar radiation intensity sequence, wind speed sequence, current and historical output power curves, remaining energy storage capacity, and equipment operating efficiency indicators. The feature extraction process includes window aggregation, such as instantaneous values, moving averages, maximum / minimum, difference and ratio calculations, such as temperature difference, output ratio, and availability flag generation, such as equipment controllability flags, shading / fault flags, and simple engineering mapping, such as obtaining the heat pump COP by temperature lookup table. The first and second feature information are output as structured arrays, facilitating subsequent time-series processing and model input.

[0032] The first feature information is subjected to time-series expansion, lag term construction and normalization to obtain the first model input feature adapted to heat load prediction; Specifically, for the first feature information, it is first unfolded temporally according to the prediction step size, that is, the temporal context of the past N steps and the future steps to be predicted is organized into a fixed shape. Then, lag terms and window statistics are constructed to explicitly introduce short-term dynamics and inertia into the input space. Subsequently, the numerical features are normalized, and categorical or periodic factors are encoded, finally forming the first model input features that match the shape of the model input layer. This process preserves time-dependent information and eliminates scale differences, which facilitates the first machine learning model to learn the temporal pattern of heat load and improves prediction stability.

[0033] The second feature information is smoothed, trend extracted, and seasonal factor labeled to obtain the second model input features adapted to capacity forecasting. Specifically, regarding the second feature information, the original sequence is first smoothed, such as by moving average or exponential smoothing, to suppress instantaneous noise. Then, through simple trend extraction, such as short-time linear regression or low-pass filtering, persistent changes are identified, and obvious seasonal / intra-day cycles are labeled, such as hourly factors, weekday / holiday factors, and monthly factors. At the same time, rated curves or temperature-efficiency lookup table results are added according to equipment characteristics. After the above processing and the necessary normalization and missing labeling, the second model input feature for capacity forecasting is obtained. This feature can reflect both environment-driven capacity trends and maintain robustness to short-term disturbances.

[0034] The first model input features are input into the first machine learning model for inference processing to obtain the target heat load prediction result; Specifically, the input features of the constructed first model are fed into the pre-trained first machine learning model in batches or streaming for inference processing. The model output is the target heat load prediction result within a preset time period, which may include point estimates for each prediction period and necessary confidence / uncertainty indices (e.g., multiple scenarios or confidence intervals). Implementation can employ parallel inference, model ensemble, or quantization for acceleration. After inference, physical boundary checks are performed on the output (e.g., not exceeding the upper limit of heat loss), and the results are calibrated online with historical observations to correct system biases. The purpose of this step is to proactively quantify existing environmental and demand information into heat demand curves that can be directly used for scheduling, thereby providing an accurate basis for group control decisions.

[0035] The second model's input features are input into the pre-trained first machine learning model for inference processing to obtain the production capacity prediction result.

[0036] Specifically, the input features of the second model are fed into a first machine learning model with the same or shared structure for inference processing. The model output is the capacity prediction results for various renewable energy devices or groups of devices within a preset time period. During inference, the post-mapping of equipment operating constraints should be considered, such as cropping the predicted values ​​to the instantaneous available power range of the equipment, and fault tolerance processing, such as backoff factors, should be implemented for short-term sudden failures or grid connection restrictions. The capacity prediction results, used in conjunction with the heat load prediction results, can form a supply and demand comparison view, providing a reliable supply-side curve for subsequent multi-objective group control optimization, thereby improving the utilization rate of renewable energy and reducing scheduling risks.

[0037] Preferably, the step of extracting first feature information for heat load prediction and second feature information for production capacity prediction based on the target environmental data, target heating demand information, and target operating status parameters includes: Based on the target environmental data, time series sampling and moving average processing are performed on the environmental temperature, environmental humidity, wind speed and solar radiation intensity to obtain a subset of environmental features characterizing external climate conditions. Specifically, based on the target environmental data, the original observation values ​​of environmental temperature, environmental humidity, wind speed, and solar radiation intensity are first read from the time-series database at a preset sampling frequency and timestamped. Then, a moving average or exponential smoothing is applied to each indicator to suppress instantaneous noise, and instantaneous values, short-term moving averages, and long-term moving averages are calculated simultaneously as parallel features. In practice, missing values ​​need to be processed (nearest neighbor interpolation or setting to the last valid value and labeling), obvious outliers need to be removed, and quality labels need to be recorded. The final output subset of environmental features includes both instantaneous quantities reflecting short-term disturbances and smoothed quantities reflecting trends and periods, which facilitates the subsequent construction of climate-sensitive prediction inputs and improves the robustness of heat load and production capacity predictions.

[0038] Based on the target heating demand information, the target set temperature, expected heating time and historical heating records are marked by time period and heating habits are summarized to obtain a subset of demand features that characterize user demand. Specifically, the target set temperature and desired heating time for users are labeled by intraday time period (e.g., morning / daytime / nighttime) and weekly category (weekday / rest day). The typical set temperature, common heating duration, frequency, and manual coverage rate for each time period are statistically analyzed from historical heating records. This allows for the summarization of heating habits (e.g., common set values, preferred temperature ranges, common early / late heating behaviors). The implementation methods include windowed statistics (calculating the mean, variance, and probability of occurrence) and weighted merging of the most recent behaviors. The output subset of demand features provides personalized input for the model, enabling prediction and scheduling to take into account both the group's optimal needs and individual comfort.

[0039] Based on the target operating state parameters, the equipment start-up and shutdown status, output power curve and operating efficiency index are screened and normalized to obtain a subset of operating features that characterize the equipment's operating characteristics. Specifically, the system extracts start-stop status signals, instantaneous output power, and estimated or measured operating efficiency (such as heat pump COP and inverter efficiency) from the equipment controller or inverter. It then performs noise reduction and smoothing (such as short-time moving average) on the original curves, eliminates obvious anomalies, and normalizes the power values ​​to relative load ratios based on the equipment's rated parameters. At the same time, it calculates key statistics (such as average output, peak value, ramp rate, and start-stop frequency). The output subset of operating characteristics reflects both the current available capacity of the equipment and the dynamic characteristics of the equipment, providing directly usable features for capacity assessment and constraint modeling.

[0040] The environmental feature subset, demand feature subset and operation feature subset are correlated and analyzed to extract the first feature information, which includes the temperature sequence outside the area to be heated, the target temperature sequence inside the area to be heated and the historical load difference sequence. Specifically, by performing correlation processing on the three subsets of environmental features, demand features, and operational features in the time series dimension (e.g., calculating the delay correlation across sequences, covariant statistics within the time window, and temperature difference-load response coefficient), the core quantities that can drive changes in heat load are directly extracted: the fused and weighted temperature sequence outside the heating area is used as the external driving input, the target temperature sequence within the heating area merged at the user end is used as the demand-side target input, and the historical load difference sequence is calculated by the difference between the historical measured load and the baseline load to reflect abnormal heating or behavioral deviations; in practice, these sequences are subjected to consistent time-series alignment, differencing, and standardization processing, and the resulting first feature information can be directly used in the heat load prediction model to improve the ability to characterize the coupling relationship between environment, demand, and operation.

[0041] The environmental feature subset and the operational feature subset are correlated and analyzed to extract the second feature information, which includes solar radiation intensity sequence, wind speed sequence and equipment operation efficiency index sequence.

[0042] Specifically, focusing on the supply side, this study maps the irradiance and wind speed values ​​in the environmental feature subset to the historical output and efficiency indicators of equipment in the operational feature subset. For example, it uses empirical power curves or lookup tables to map solar radiation intensity to the expected output of photovoltaic or solar thermal power, and wind speed to wind power output through wind turbine power curves. At the same time, it extracts the equipment operation efficiency index sequence under the same environmental conditions. In practice, the mapping results are smoothed and normalized over time. The output second feature information includes the solar radiation intensity sequence, wind speed sequence, and equipment operation efficiency index sequence, which directly characterize the energy supply potential of renewable energy equipment groups in each forecast period, providing a clear quantitative basis for capacity forecasting and subsequent group control constraint modeling.

[0043] Preferably, the pre-training process of the first machine learning model includes: A training sample set is constructed based on pre-collected historical multi-source environmental data, historical heating demand information, and corresponding historical operating status parameters of renewable energy equipment; Specifically, the historical multi-source environmental data may include long-term recorded outdoor / indoor temperature, humidity, wind speed, solar radiation, etc.; historical heating demand information includes past user-set temperatures, heating periods, and actual heating consumption; and historical operating status parameters include equipment start / stop logs, real-time output, and efficiency records. The aim is to pair these three types of historical records according to time cues to create training samples representing real-world operating scenarios, providing material for the model to learn the supply-demand mapping relationship. In practice, the sample time resolution must first be confirmed. Then, based on a unified time scale, environmental, demand, and equipment data are sliced, aggregated by window, and labeled with corresponding target values. Each slice is encapsulated as a training sample and stored in a training sample library, forming a rich training sample set covering multiple operating conditions, seasons, and user behaviors, thereby improving the model's generalization ability and robustness.

[0044] The historical multi-source environmental data is denoised, missing values ​​are imputed, and timestamps are aligned to obtain the environmental time series for training. Specifically, denoising can employ median filtering, moving average, or outlier detection and removal to eliminate sensor jitter and instantaneous errors; missing value imputation can use linear interpolation, forward imputation, or spatial interpolation based on adjacent stations to restore continuous time series for time series modeling; timestamp alignment involves resampling asynchronously sampled multi-source environmental data to a unified time grid and correcting clock drift. In implementation, anomaly detection and quality label generation are performed first in the data pipeline, followed by resampling and interpolation at the target resolution. The final output is a training environmental time series with quality labels, ensuring the temporal consistency and quality of the environmental input and reducing training bias caused by time misalignment and noise.

[0045] The historical heating demand information is formatted, normalized, and labeled with user habit features to obtain the demand time series for training. Specifically, formatting unifies user input from various sources (thermostat settings, APP commands, BMS instructions) into structured fields; normalization converts different dimensions (temperature, duration, frequency) into comparable scales; user habit feature annotation includes extracting typical set temperatures, commonly used heating periods, and manual overlay frequency to represent user preferences. The implementation process involves: extracting historical records, aggregating time periods (by hour / day), calculating statistics (mean, variance, probability of occurrence), encoding categorical variables, and outputting the demand time series for training. This preserves user behavior characteristics, enabling the model to capture personalized needs while learning heat load patterns, thus improving the adaptability of predictions in individual and group scenarios.

[0046] The historical operating state parameters are smoothed, standardized, and efficiency factors are calculated to obtain the equipment operating time series for training. Specifically, smoothing is used to suppress short-term fluctuations at equipment measurement points (e.g., using moving average or exponential smoothing); standardization normalizes indicators such as power to relative load ratios based on rated values; efficiency factor calculation includes estimating the COP of heat pumps or the efficiency curves of inverters / generators based on temperature-power relationships. First, start-stop pulses are de-jittered, then instantaneous output is corrected using rated parameters, and statistics (average output, peak-to-valley ratio, start-stop frequency) are calculated. Finally, the equipment operating time series with efficiency factors is output. By generating training inputs that accurately characterize the equipment's capabilities and dynamic characteristics, the model can learn how production capacity changes with environmental and operating conditions.

[0047] Based on the environmental time series, demand time series, and equipment operation time series, first training feature information for heat load prediction and second training feature information for capacity prediction are extracted respectively. Specifically, the first training feature information consists of the external temperature sequence of the area to be heated, the target temperature sequence within the area to be heated, the historical load baseline, the average temperature difference within the time window, and the time / day cycle factor; the second training feature information consists of the solar radiation sequence, wind speed sequence, relative equipment output sequence, energy storage charge rate, and efficiency factor. The implementation involves performing sliding window aggregation (mean, maximum, minimum, variance) on the three types of time series—environmental time series, demand time series, and equipment operation time series—constructing lag terms and difference terms, and periodically encoding the time features (hour / week / holiday), outputting two types of structured feature vectors as the first training feature information for heat load prediction and the second training feature information for capacity prediction; transforming the original time series into training features with physical meaning and easy learning, improving the model's ability to capture short-term dynamic and periodic changes.

[0048] The first training feature information is matched with the actual historical heat load observations to construct the first training data pair; Specifically, the aforementioned first training feature information is paired with the synchronously obtained historical heat load observations according to the same time window. If necessary, the target value is smoothed or anomaly corrected to reduce label noise. In implementation, the target variable definition (instantaneous power, time period energy or peak value, etc.) and alignment strategy (the feature window corresponds to the label of the future prediction step) need to be clearly defined and stored in an iteratively readable data pair format (feature matrix + label vector) to obtain high-quality supervised training samples, so that the heat load sub-model can learn the causal mapping from environment and demand to actual load.

[0049] The second training feature information is matched with the actual observed values ​​of equipment capacity to construct a second training data pair; Specifically, the second training feature information needs to correspond one-to-one with the actual production capacity observation of the equipment (such as the actual available output or net energy supply of photovoltaic / wind turbine / heat pump) according to the time window. If the observation is affected by faults or grid connection restrictions, it needs to be labeled and masked in the training. The implementation process includes determining the definition of production capacity target (instantaneous available power or future cumulative energy), aggregating or decomposing the production capacity of a single unit or group, and generating feature-label pairs for training, providing real supply-side mapping data for the production capacity prediction sub-model, and improving the modeling accuracy of the volatility of renewable resources.

[0050] The first training data pair and the second training data pair are respectively input into the first machine learning model for supervised training until the prediction error of the model meets the preset convergence condition, thus obtaining the first machine learning model after training.

[0051] Specifically, the training data is divided into training / validation / test sets. A suitable loss function is selected (e.g., MSE / MAE can be used for heat load, and quantile loss can be added to capacity to address uncertainty). Iterative training is performed using mini-batch gradient descent or other optimizers, and overfitting is controlled through early stopping, learning rate scheduling, and cross-validation. During training, the model's performance on the validation set is evaluated, and parameters are adjusted or features are replaced as needed until the prediction error meets the preset convergence condition (e.g., the validation set RMSE is less than a threshold or the error no longer decreases significantly). Finally, the first machine learning model and calibration parameters are saved. By obtaining a first machine learning model that can stably and accurately perform heat load and capacity inference, reliable prediction support is provided for subsequent real-time inference and swarm control optimization.

[0052] Preferably, please refer to Figure 2 The step of determining the group control strategy for the renewable energy equipment group based on a preset second machine learning model, combined with the target heat load prediction result and the capacity prediction result, includes: Based on the target heat load prediction results, obtain the heat demand information of the area to be heated in each prediction period; Specifically, the target heat load forecast results (usually represented by power or energy curves for each forecast period) are first summarized and converted into heat demand (e.g., kW or kWh / per period) according to the required time resolution. At the same time, the forecast uncertainty (confidence interval or multiple scenarios) is added as an uncertainty description on the demand side. Secondly, the demand is zoned and aggregated (allocated by building / floor / room) and preheating, heat loss, and internal heat source correction factors are considered to obtain structured heat demand information for each forecast period (including center value, upper / lower bounds, and regional allocation). This process can be implemented at the edge or in the cloud through data mapping rules and energy balance correction. The purpose is to provide directly usable quantitative demand curves for subsequent scheduling. Its beneficial effect is to transform the model output into an executable demand input, reduce the risk of misallocation, and support fine-grained scheduling by zone allocation.

[0053] Based on the capacity forecast results, obtain the energy availability information of the renewable energy equipment group in each forecast period; Specifically, the capacity forecast results (the theoretical output of each device or group of devices according to time periods) are processed into available energy supply information for scheduling. This includes grid connection / inverter quota reduction for the capacity of individual devices, consideration of grid connection constraints or output priorities, and incorporation of the charging and discharging capacity of energy storage / thermal storage devices into the available energy supply summary. At the same time, the available upper limit, adjustable range, and reserve capacity (considering failure retirement rate or maintenance plan) for each forecast period are calculated, and short-term fluctuations are processed using smoothing or conservative backoff factors to form a robust energy supply curve. In practice, this is achieved through table lookup (equipment power curve), rule-based backoff, and real-time communication backfilling. The result is available energy supply information that facilitates constraint modeling and optimization input. Its beneficial effect is to provide a realistic and executable supply-side capacity curve and avoid generating unrealistic scheduling schemes.

[0054] Obtain a pre-established group control optimization objective function, wherein the objective function includes at least a heating comfort objective, an energy utilization objective, and an operating cost objective; Specifically, each sub-objective is clearly defined and quantified: the heating comfort objective is defined as the weighted sum of squares of heating temperature deviations or the occupancy violation rate; the energy utilization rate objective is defined as the proportion of supply and demand directly met by renewable energy or the renewable energy utilization rate; and the operating cost objective is defined as the weighted sum of equipment fuel consumption, electricity purchase cost, start-up and shutdown cost, and maintenance amortization. Then, the weights of each objective are selected or dynamically determined according to the strategy, or retained as a multi-objective form to obtain the Pareto solution set. At the same time, penalty terms (such as over-temperature penalty, frequent start-up and shutdown penalty) and priority rules (user priority / safety priority) in the objective function are defined. In implementation, these mathematical expressions are encapsulated into an objective function module that can be called by the optimizer (supporting weighted sum, constraint optimization, or multi-objective evolution solution). During scheduling, comfort, utilization, and cost can be clearly weighed, which facilitates obtaining an implementable group control objective that meets the strategy requirements.

[0055] Based on the heat demand information and the available energy information, construct supply and demand balance constraints; Specifically, demand and supply are constrained in the form of equations for each forecast period; at the same time, energy state equations (such as the time-series update formula of energy storage SOC) are added to the time series continuity, and minimum reserve ratio or safety margin constraints can be introduced to deal with forecast uncertainty; in practice, these expressions are discretized into linear or nonlinear constraints (depending on whether efficiency curves are considered), and time-series constraint sets are generated for the optimizer to ensure that the group control strategy meets the balance requirements at the energy level, fundamentally avoiding supply and demand mismatch and loss of comfort.

[0056] Based on the preset start-stop logic, output power limit, and operating efficiency boundary of the renewable energy equipment group, construct the equipment operation constraints; Specifically, executable constraints are established for each type of equipment, including: binary start / stop logic and minimum start / stop duration, upper and lower limits of output power, maximum ramp / ramp rate, efficiency boundary (which can be used as a nonlinear constraint or approximated as a piecewise linear function), maintenance / downtime window, and energy storage charge / discharge rate and SOC upper and lower limits. During implementation, these physical and safety constraints are formatted into a constraint matrix that the optimizer can recognize (using mixed integer form for binary variables, and piecewise approximation or pre-calculated interpolation tables for nonlinear efficiency), and priorities and default costs are marked in the constraint set (for soft constraint handling). This ensures that the generated group control strategy can be directly executed at the equipment level without impairing equipment lifespan or violating safety rules.

[0057] The supply and demand balance constraints and equipment operation constraints are input into the second machine learning model, and combined with the group control optimization objective function, multi-objective optimization is performed to obtain the solution result; Specifically, the constraint set and objective function are first converted into an input format usable by a second machine learning model: if a learning-based optimization agent is used, the constraints are represented as states or embedded into the objective through penalty functions; if a hybrid optimizer is used, the constraints are passed to the solver. Then, a multi-objective optimization solution process is executed: a set of candidate policies is generated using weighted sum / hierarchical optimization or evolutionary multi-objective algorithms (such as NSGA), or policy parameters are directly output using a trained policy network / reinforcement learning agent. After each iteration or generation of candidate solutions, unqualified solutions are eliminated through rapid simulation or constraint checks, and the objective function value is evaluated. This process is repeated until convergence or the computational budget is reached, outputting a solution that satisfies the constraints and is dominant on multiple objectives (which may include a set of Pareto solutions). This systematic, multi-objective approach seeks the optimal trade-off between comfort, utilization, and cost, and provides an executable set of swarm control policy candidates within an acceptable computational time.

[0058] Based on the solution results, the start / stop commands and power allocation schemes for the device group in each prediction period are obtained, thus obtaining the group control strategy.

[0059] Specifically, the optimized solution is transformed into a concrete group control strategy: the start-stop decisions, time-series power curves of each device, and energy storage charging and discharging plans from the candidate solutions are solidified into timestamped control plans, and backup schemes and rollback rules are generated for critical periods; before issuance, the strategy is validated for feasibility (device constraints and supply-demand balance are checked again), and start-stop instructions and power allocation schemes are issued to the control terminals of each device according to priority, while recording the instruction version and execution window for auditing; during operation, execution deviations are monitored in real time, and rescheduling or local compensation strategies are triggered when necessary. The abstract solution of multi-objective optimization is implemented into an executable, monitorable, and redundant group control strategy, thereby achieving heating operation that satisfies both comfort and maximizes the utilization and economy of renewable energy in each forecast period.

[0060] Preferably, the step of inputting the supply-demand balance constraints and equipment operation constraints into the second machine learning model, and combining them with the group control optimization objective function to perform multi-objective optimization and obtain the solution results includes: Using the supply-demand balance constraints and the equipment operation constraints as input constraints, and combining them with the group control optimization objective function, a group control optimization problem with multiple constraints is constructed. Specifically, the supply-demand balance constraints are first formalized as a dynamic balance between heating demand and renewable energy supply to ensure that energy supply can cover the corresponding heat demand during each forecast period. Furthermore, the equipment operation constraints are transformed into limitations on operational boundaries such as equipment start-up and shutdown logic, output power range, power change rate, and minimum start-up and shutdown duration to ensure equipment operational safety and physical feasibility. Subsequently, these constraints are combined with a pre-defined group control optimization objective function, which covers key indicators such as heating comfort, energy utilization rate, and operating cost, thus forming a group control optimization problem with multiple constraints, providing a foundation for subsequent optimization calculations.

[0061] The second machine learning model is used to iteratively solve the group control optimization problem. In each iteration, the start / stop commands and power allocation parameters of the device group are adjusted to obtain multiple candidate group control strategies. Specifically, in the solution process, the second machine learning model plays a role in rapidly approximating and dynamically adjusting complex optimization problems. Specifically, in each iteration, the second machine learning model takes into account the current group control state, constraint satisfaction, and the evaluation result of the objective function, and then outputs adjustment suggestions for device start / stop commands and power allocation schemes. Through multiple rounds of iterative calculations, the model continuously refines candidate schemes, gradually approximating a set of feasible and high-performance solutions. After multiple iterations, several candidate group control strategies satisfying different optimization preferences can be generated, thus providing diverse choices for subsequent selection.

[0062] The heating comfort, energy efficiency and operating cost of each candidate group control strategy are comprehensively scored, and the candidate group control strategies are ranked according to the scoring results to obtain the ranking results; Specifically, this step requires a quantitative evaluation of each candidate group control strategy from multiple dimensions. First, heating comfort indicators are calculated, such as whether the area to be heated remains within the set temperature range during each predicted period. Second, energy utilization indicators are calculated, such as the proportion of renewable energy used in the heating process. Finally, operating cost indicators are calculated, including energy consumption costs and maintenance costs associated with equipment start-up and shutdown. These indicators are then standardized and weighted, or a multi-objective ranking method is used to rank the candidate group control strategies according to their overall merits, thus obtaining a clear ranking result.

[0063] Based on the ranking results, the optimal candidate group control strategy is selected as the solution result.

[0064] Specifically, the optimal solution is determined based on the ranking results and becomes the final solution. The selected group control strategy includes equipment start-up and shutdown commands and power allocation schemes for each forecast period, ensuring a reasonable balance between comfort, energy utilization, and cost. Before being issued for execution, the solution undergoes feasibility verification to ensure complete compatibility with equipment operating logic, and is then issued as a group control command to the equipment control terminal for implementation. This not only improves the comprehensive utilization efficiency of renewable energy but also reduces overall operating costs while ensuring heating comfort.

[0065] Preferably, the pre-training process of the second machine learning model includes: Based on the pre-collected historical target heat load prediction results and historical production capacity prediction results, combined with historical heating comfort index, energy utilization rate index and operating cost data, a multi-objective training sample set is constructed. Specifically, the historical target heat load forecast results and historical production capacity forecast results are derived from the historical operation forecast module or archived records, respectively. Historical heating comfort indicators refer to recorded temperature deviations, violation periods, or user complaint statistics. Energy utilization rate indicators refer to the proportion of direct heating from renewable energy sources, etc. Operating cost data refers to energy consumption costs and equipment operation and maintenance costs. The purpose is to combine supply and demand forecasts with operational consequences into a sample that can be learned so that the second machine learning model can learn the trade-off relationship between the objectives. In practice, these historical quantities need to be summarized at a uniform time resolution, scene labels (such as seasons, holidays, extreme weather events) should be injected, and abnormal samples should be removed or labeled. Finally, a training sample set covering multiple operating conditions and multiple objectives should be formed to provide subsequent models with rich historical experience data that can reflect the trade-off between comfort, utilization rate, and cost.

[0066] The historical heat load prediction results and historical production capacity prediction results are normalized and time-series aligned to obtain the input feature sequence for training. Specifically, firstly, the historical target heat load prediction results and historical production capacity prediction results are resampled at predetermined time grids and timestamped to achieve time series alignment. Secondly, the numerical range and units are normalized to eliminate the influence between different units. At the same time, confidence or error estimates are added as additional input features for each time period. In terms of implementation, missing value completion strategies, smoothing or denoising methods are adopted, and various predictions are organized into sequence samples with a uniform window width. The output training input feature sequence has both temporal consistency and numerical comparability, which facilitates the second machine learning model to accurately capture the time series patterns of supply and demand changes during training.

[0067] The historical heating comfort index, energy utilization rate index and operating cost data are standardized and weighted to obtain a multi-objective output sequence for training. Specifically, historical heating comfort indicators, energy utilization rate indicators, and operating cost data are processed using a unified standardized method to make each indicator comparable. Weights are assigned to each indicator based on strategy or operational preferences to reflect priority differences. In practice, the calculation criteria for each indicator need to be defined, and abnormal or missing target values ​​need to be labeled or masked. If necessary, long-term and short-term objectives are encoded as hierarchical outputs. The resulting multi-objective output sequence retains the independence of each objective while representing the relative importance between objectives through weight information, thus providing clear learning labels for multi-objective supervised training.

[0068] By mapping the input feature sequence to the multi-target output sequence, training data pairs are constructed. Specifically, the normalized and aligned input feature sequences are precisely paired with the standardized and weighted multi-objective output sequences within the same time window. A large number of training sample pairs are generated using a sliding window or fixed window strategy, and samples containing anomalies or low confidence are screened out or labeled to prevent training bias. In the process, training / validation / test sets need to be divided, and necessary data augmentation or undersampling is performed on the samples to balance the distribution of different scenarios. Finally, a set of training data pairs that can be directly read by the model is obtained, which provides a structured data foundation for subsequent multi-objective supervised training.

[0069] The training data is input into the second machine learning model, and a multi-objective optimization algorithm is used for supervised training. During the training process, the heating comfort target, energy utilization target and operating cost target are balanced through dynamic weight allocation. Specifically, the training data is used to supervise learning the input of the second machine learning model. During training, optimization strategies capable of handling multiple objectives are employed, such as weighted loss, dynamic weight adjustment, or a multi-task learning framework, to balance the various objectives. The dynamic weight allocation method automatically adjusts the weights by monitoring the performance of each objective on the validation set to prevent any one objective from dominating the training process. In terms of implementation, this includes selecting a suitable model structure, configuring the optimizer and learning rate strategy, using mini-batch training, and periodically evaluating the indicators of each objective during training to drive weight adjustment. The training result is a candidate model that performs well in the multi-objective space.

[0070] The training continues iteratively until the combined error of the second machine learning model on the validation set meets the preset convergence condition, thus obtaining the trained second machine learning model.

[0071] Specifically, the training process is conducted iteratively, continuously monitoring the model's overall error on the validation set. This overall error is calculated based on a predefined multi-objective aggregation criterion and used to determine convergence. The model is considered to have converged when the overall error reaches a set threshold or no longer improves significantly within several iterations. Implementation details include an early stopping mechanism, hyperparameter tuning, and cross-validation to prevent overfitting. After training, the model weights and a performance snapshot from the validation process are saved, and a training report is generated. The resulting trained second machine learning model has the decision-making ability to simultaneously consider heating comfort, energy utilization, and operating costs in group control optimization, and can serve as the basis for offline or online strategy generation.

[0072] Example 2 In addition, combined Figure 1 The machine learning-based renewable energy heating group control method described in Embodiment 1 of the present invention can be implemented by a machine learning-based renewable energy heating group control system. Figure 3 A schematic diagram of the hardware structure of the renewable energy heating group control system based on machine learning provided in Embodiment 2 of the present invention is shown.

[0073] A machine learning-based renewable energy heating group control system may include a processor and a memory storing computer program instructions.

[0074] Specifically, the processor may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement embodiments of the present invention.

[0075] The memory may include a large-capacity storage device for data or instructions. For example, and not limitingly, the memory may include a hard disk drive (HDD), a floppy disk drive, flash memory, an optical disk drive, a magneto-optical disk drive, magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, the memory may include removable or non-removable (or fixed) media. Where appropriate, the memory may be internal or external to a data processing device. In a particular embodiment, the memory is a non-volatile solid-state memory. In a particular embodiment, the memory includes a read-only memory (ROM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically rewritable ROM (EAROM), or flash memory, or a combination of two or more of these.

[0076] The processor reads and executes computer program instructions stored in memory to implement any of the machine learning-based renewable energy heating group control methods in the above embodiments.

[0077] In one example, a machine learning-based renewable energy heating group control system may also include a communication interface and a bus. For example, Figure 3 As shown, the processor, memory, and communication interface are connected via a bus and communicate with each other.

[0078] The communication interface is mainly used to enable communication between various modules, devices, units and / or equipment in the embodiments of the present invention.

[0079] A bus, including hardware, software, or both, couples components of the device together. For example, and not limitingly, a bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, a bus may include one or more buses. While specific buses are described and illustrated in embodiments of the invention, the invention contemplates any suitable bus or interconnect.

[0080] In summary, the embodiments of the present invention provide a method and system for group control of renewable energy heating based on machine learning.

[0081] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0082] The functional blocks shown in the above-described structural diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the required tasks. The programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0083] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data shall comply with the relevant laws, regulations and standards of the relevant locality, and corresponding operation entry points shall be provided for the user to choose to authorize or refuse.

[0084] It should also be noted that the exemplary embodiments mentioned in this invention describe methods or systems based on a series of steps or apparatus. However, this invention is not limited to the order of the steps described above; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0085] The above description is merely a specific embodiment of the present invention. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the protection scope of the present invention.

Claims

1. A machine learning-based method for group control of renewable energy heating, characterized in that, The method includes: Acquire multi-source environmental data of the area to be heated, user heating demand information, and operating status parameters of renewable energy equipment groups used for heating; Using a pre-trained first machine learning model, the multi-source environmental data, the heating demand information, and the operating status parameters are processed to obtain the target heat load prediction results for the area to be heated and the capacity prediction results for the renewable energy equipment group within a preset time period. Based on the preset second machine learning model, and combining the target heat load prediction results and the capacity prediction results, the group control strategy for the renewable energy equipment group is determined; According to the group control strategy, the renewable energy equipment group is controlled to complete the heating of the area to be heated.

2. The machine learning-based renewable energy heating group control method according to claim 1, characterized in that, The acquisition of multi-source environmental data of the area to be heated, user heating demand information, and operating status parameters of the renewable energy equipment group used for heating includes: By deploying environmental sensors inside and outside the area to be heated, the first environmental parameters of multiple preset local areas within the area to be heated and the second environmental parameters within a preset area outside the area to be heated are collected. Based on the preset weights corresponding to each preset local region, the first environmental parameters are weighted and fused to obtain the multi-source environmental parameters within the area to be heated. The second environmental parameter and the multi-source environmental parameter are subjected to time-series synchronization processing to obtain the multi-source environmental data; The start-stop status, output power, and operating efficiency data of the renewable energy equipment group are collected to obtain the operating status parameters; Based on the target set temperature and desired heating time input by the user, and combined with the user's historical heating preference information, the heating demand information is obtained.

3. The renewable energy heating group control method based on machine learning according to claim 1, characterized in that, The process of using a pre-trained first machine learning model to process the multi-source environmental data, the heating demand information, and the operating status parameters to obtain the target heat load prediction results for the area to be heated and the capacity prediction results for the renewable energy equipment group within a preset time period includes: The multi-source environmental data, the heating demand information, and the operating status parameters are preprocessed to obtain preprocessed target environmental data, target heating demand information, and target operating status parameters. Based on the target environmental data, target heating demand information, and target operating status parameters, extract the first feature information for heat load prediction and the second feature information for production capacity prediction. The first feature information is subjected to time-series expansion, lag term construction and normalization to obtain the first model input feature adapted to heat load prediction; The second feature information is smoothed, trend extracted, and seasonal factor labeled to obtain the second model input features adapted to capacity forecasting. The first model input features are input into the first machine learning model for inference processing to obtain the target heat load prediction result; The second model's input features are input into the pre-trained first machine learning model for inference processing to obtain the production capacity prediction result.

4. The machine learning-based renewable energy heating group control method according to claim 3, characterized in that, The step of extracting first feature information for heat load prediction and second feature information for production capacity prediction based on the target environmental data, target heating demand information, and target operating status parameters includes: Based on the target environmental data, time series sampling and moving average processing are performed on the environmental temperature, environmental humidity, wind speed and solar radiation intensity to obtain a subset of environmental features characterizing external climate conditions. Based on the target heating demand information, the target set temperature, expected heating time and historical heating records are marked by time period and heating habits are summarized to obtain a subset of demand features that characterize user demand. Based on the target operating state parameters, the equipment start-up and shutdown status, output power curve and operating efficiency index are screened and normalized to obtain a subset of operating features that characterize the equipment's operating characteristics. The environmental feature subset, demand feature subset and operation feature subset are correlated and analyzed to extract the first feature information, which includes the temperature sequence outside the area to be heated, the target temperature sequence inside the area to be heated and the historical load difference sequence. The environmental feature subset and the operational feature subset are correlated and analyzed to extract the second feature information, which includes solar radiation intensity sequence, wind speed sequence and equipment operation efficiency index sequence.

5. The renewable energy heating group control method based on machine learning according to claim 1, characterized in that, The pre-training process of the first machine learning model includes: A training sample set is constructed based on pre-collected historical multi-source environmental data, historical heating demand information, and corresponding historical operating status parameters of renewable energy equipment; The historical multi-source environmental data is denoised, missing values ​​are imputed, and timestamps are aligned to obtain the environmental time series for training. The historical heating demand information is formatted, normalized, and labeled with user habit features to obtain the demand time series for training. The historical operating state parameters are smoothed, standardized, and efficiency factors are calculated to obtain the equipment operating time series for training. Based on the environmental time series, demand time series, and equipment operation time series, first training feature information for heat load prediction and second training feature information for capacity prediction are extracted respectively. The first training feature information is matched with the actual historical heat load observations to construct the first training data pair; The second training feature information is matched with the actual observed values ​​of equipment capacity to construct a second training data pair; The first training data pair and the second training data pair are respectively input into the first machine learning model for supervised training until the prediction error of the model meets the preset convergence condition, thus obtaining the first machine learning model after training.

6. The machine learning-based renewable energy heating group control method according to any one of claims 1-5, characterized in that, The step of determining the group control strategy for the renewable energy equipment group based on a preset second machine learning model, combined with the target heat load prediction result and the capacity prediction result, includes: Based on the target heat load prediction results, obtain the heat demand information of the area to be heated in each prediction period; Based on the capacity forecast results, obtain the energy availability information of the renewable energy equipment group in each forecast period; Obtain a pre-established group control optimization objective function, wherein the objective function includes at least a heating comfort objective, an energy utilization objective, and an operating cost objective; Based on the heat demand information and the available energy information, construct supply and demand balance constraints; Based on the preset start-stop logic, output power limit, and operating efficiency boundary of the renewable energy equipment group, construct the equipment operation constraints; The supply and demand balance constraints and equipment operation constraints are input into the second machine learning model, and combined with the group control optimization objective function, multi-objective optimization is performed to obtain the solution result; Based on the solution results, the start / stop commands and power allocation schemes for the device group in each prediction period are obtained, thus obtaining the group control strategy.

7. The machine learning-based renewable energy heating group control method according to claim 6, characterized in that, The process involves inputting the supply-demand balance constraints and equipment operation constraints into the second machine learning model, and combining this with the group control optimization objective function to perform multi-objective optimization, yielding the following results: Using the supply-demand balance constraints and the equipment operation constraints as input constraints, and combining them with the group control optimization objective function, a group control optimization problem with multiple constraints is constructed. The second machine learning model is used to iteratively solve the group control optimization problem. In each iteration, the start / stop commands and power allocation parameters of the device group are adjusted to obtain multiple candidate group control strategies. The heating comfort, energy efficiency and operating cost of each candidate group control strategy are comprehensively scored, and the candidate group control strategies are ranked according to the scoring results to obtain the ranking results; Based on the ranking results, the optimal candidate group control strategy is selected as the solution result.

8. The renewable energy heating group control method based on machine learning according to claim 1, characterized in that, The pre-training process of the second machine learning model includes: Based on the pre-collected historical target heat load prediction results and historical production capacity prediction results, combined with historical heating comfort index, energy utilization rate index and operating cost data, a multi-objective training sample set is constructed. The historical heat load prediction results and historical production capacity prediction results are normalized and time-series aligned to obtain the input feature sequence for training. The historical heating comfort index, energy utilization rate index and operating cost data are standardized and weighted to obtain a multi-objective output sequence for training. By mapping the input feature sequence to the multi-target output sequence, training data pairs are constructed. The training data is input into the second machine learning model, and a multi-objective optimization algorithm is used for supervised training. During the training process, the heating comfort target, energy utilization target and operating cost target are balanced through dynamic weight allocation. The training continues iteratively until the combined error of the second machine learning model on the validation set meets the preset convergence condition, thus obtaining the trained second machine learning model.

9. A machine learning-based renewable energy heating group control system, characterized in that, include: At least one processor, at least one memory, and computer program instructions stored in the memory, which, when executed by the processor, implement the method as described in any one of claims 1-8.

Images