A building group heat load prediction method based on working condition self-adaptation and causal residual fusion

Abstract

The application relates to a building group heat load prediction method based on working condition self-adaptation and causal residual fusion, which comprises the following steps of data preparation and consistency processing, causal safety feature construction, interpretable baseline modeling, error learner training, self-adaptive synthesis under the heating working condition, rolling correction and parameter updating, and result quality guarantee. The application determines universal input parameters from a basic heat transfer equation, considers room temperature characteristic values representing user demand and a composite representation framework facing time sequences, comprehensively depicts representative quantities of user demand and the effect of meteorological cumulative effect on the load, and explicitly includes the influence of operation uncertainty; multiple features are used to simulate corresponding thermal characteristics and uncertainty disturbances, so that the model has interpretability; model parameters are statistically robustly calibrated by using historical operation data, and long-term slow drift and distribution light shift are inhibited through consistency correction and rolling correction, so that the reliability and stability of long-term operation are improved.

Description

Technical Field

[0001] This invention belongs to the field of centralized heating technology, and relates to the prediction of heat load for building complexes and energy stations, particularly to a method for predicting heat load for building complexes based on adaptive operating conditions and causal residual fusion. Background Technology

[0002] Central heating systems play a crucial role in creating a comfortable living and production environment in cold regions[1]. Statistical data shows that the energy consumption of central heating and domestic hot water supply accounts for 42% of the total industrial energy consumption[2]. With the advancement of global urbanization, the demand for building energy consumption continues to rise[3,4]. The inherent thermal inertia of buildings, coupled with the large scale and complex structure of central heating systems, often leads to lag effects. This effect can cause overheating within the system, resulting in a large amount of energy waste[5,6]. Therefore, load forecasting of heating systems is crucial, as it can alleviate the impact of lag effects, ensure user thermal comfort, reduce energy consumption, and achieve efficient and energy-saving system operation[7]. Accurate heating load forecasting is the foundation for efficient system operation and precise regulation, and plays a decisive role in the optimized management of heating systems[8].

[0003] Commonly used commercial software such as TRNSYS, EnergyPlus, Dymola, Blast, DEST and eQuest are widely used in energy consumption prediction. Reinhart and Cerezo Davila[9] used TRNSYS to evaluate the energy consumption of typical urban residential buildings, and the results showed a deviation of 5% to 20%. Lei Shuyao et al.

[10] used EnergyPlus to study the influence of different thermal parameters of building envelope on residential energy consumption. Cucca

[11] et al. used TRNSYS and GenOpt to analyze the energy consumption of high-rise buildings in Xi'an, but the accuracy of the model depends on the accuracy of the input building information, which is difficult to obtain in actual applications, thus limiting the applicability of the model. However, these prediction methods often require detailed information on building geometry, envelope and building materials under different operating scenarios, which leads to a significant increase in modeling complexity and computational cost.

[0004] In recent years, with the application of IoT technology and the development of automatic control technology, the automation, informatization, and intelligence levels of centralized heating systems have been generally improved. The deep integration of information and energy has made it possible to achieve precise control of "on-demand heating." Under the dual-carbon target framework, based on the characteristics of building thermal inertia and system regulation time lag, a load model and control method centered on target energy consumption control have been developed. This model provides timely demand load forecasts and control responses within a sufficiently small deviation range based on meteorological factors and actual user needs and their time-varying nature. This is of great significance for achieving clean and efficient heating in the building sector.

[0005] Khajavi et al.

[12] combined six metaheuristic algorithms with support vector regression (SVR) for hyperparameter optimization. The model can integrate the advantages of multiple methods to improve the accuracy of load prediction and achieve precise regulation. However, as the training sample size decreases, the accuracy and stability of these methods will decrease significantly.

[0006] Liu Yubin et al.

[13] used “similar weather days + unit area heat index” as the framework, and used K1 / K2 weight and L1 / L2 two-level correction to balance “experience reference” and “current needs”. Under the difference in physical sensation such as dampness and cold, the fit was improved by regular compensation. However, it relied heavily on similarity measurement, weight setting and correction rules. The model adaptation and uncertainty handling were mainly reflected in the setting of empirical parameters and thresholds. It was a static estimation and lacked systematic handling and dynamic adaptation of thermal inertia, drift and uncertainty disturbance.

[0007] Gong Mingju et al.

[14] used “similar day + deep neural network (SD-DNNs)” as the core. First, XGBoost was used to calculate the weights. The outdoor temperature, wind force and the previous day’s load were weighted to improve the Euclidean distance to select two similar days. Then, the eight features of “similar day + predicted day” were input into a four-layer fully connected DNN. After training for 400 rounds, the short-term heat load was output and evaluated by RMSE / MAPE / MPE. This scheme abstracts the complex and non-stationary heating system into a static paradigm of “similar day + fixed black box”. The similarity measurement is prone to failure when faced with temporary scheduling and operating condition switching.

[0008] Liu Rong et al.

[15] proposed a deep LSTM end-to-end prediction of heating demand load: first, determine the input and train a "black box" model with historical samples. The input includes the weather at the next moment, holiday indication, the actual load at the previous moment and room temperature. The output is the demand load at the moment to be predicted. When there are no room temperature measurement points, the correction coefficient is set according to the empirical value of supply / return water temperature given by the operators to make empirical correction to the sample load. During the operation period, the error threshold triggers online accuracy monitoring and full window retraining (including data cleaning and standardization). However, relying on empirical parameter correction, the cross-scenario transferability is poor, the model interpretability and auditability are weak, and it is difficult to locate the source of error and make fine-grained operation and maintenance closed loop.

[0009] Zhang Lijun et al.

[16] proposed a backpropagation neural network (PCA-GA-BP) prediction model based on principal component analysis and genetic algorithm optimization, and determined eight initial characteristic indicators, including outdoor temperature, wind speed, relative humidity, primary water supply temperature, return water temperature, pipeline flow rate, hourly heat at the same time one day before, and hourly heat at the same time two days before. The actual thermal data of the residential building heating system in Kangding City were used as simulation experimental samples to compare and verify the performance of the model before and after optimization. The study found that the average absolute percentage error of the PCA-GA-BP neural network prediction model was as low as 10.287%, which was 6.636% lower than that of the traditional BP neural network prediction model, significantly improving the accuracy of heating load prediction.

[0010] Wu Hao et al.

[17] based their work on historical operation and meteorological data, first processed missing and abnormal data, and then normalized and filtered for correlation to form a feature subset. On this basis, they introduced statistical learning evaluation of building thermal insulation performance and meteorological time series prediction as auxiliary information. Finally, they used a deep time series attention model to jointly model multiple source elements and output short-term heating load prediction results, which reflected the data-driven integrated modeling and multi-source information coupling ideas.

[0011] However, in practical applications, the challenge of insufficient data is often encountered, such as the difficulties in upgrading and transforming old residential communities into smart homes, and the limited monitoring data available for the system. To ensure the establishment of robust models applicable to datasets of different sizes and to provide accurate predictions, relying solely on data-driven machine learning methods remains insufficient.

[0012] On the other hand, in heat load models, the input dimension and model performance need to be in a proper balance: too many variables will increase complexity and lengthen training and simulation time, while too few variables may sacrifice accuracy. This trade-off is crucial

[18] . Related studies have shown that task-oriented feature identification and dimensionality reduction can significantly improve the efficiency and generalization performance of the model while maintaining interpretability

[19] . It should be noted that although indoor temperature is the core state quantity of residential building heating design and operation, it is often set as an approximately constant target value in engineering practice for ease of operation control and comfort assurance. This weakens the role of its time-varying nature in load modeling to some extent.

[0013] The main purpose of centralized heating system operation and regulation is to ensure that the heating system meets users' heat demands while avoiding energy waste caused by overheating. Centralized heating systems typically include a large number of users of various types, whose heat consumption patterns differ significantly, and even users of the same type do not have exactly the same heat demand.

[0014] Indoor temperature is a key parameter in the design and operation of residential building heating systems, but it is often set to a fixed value to meet the constant needs of different users. Nik et al.

[20] evaluated the energy-saving potential of energy-saving renovation measures and pointed out that adjusting the indoor temperature can achieve significant energy savings. The needs of different user types have dynamic characteristics. Maintaining a constant indoor temperature for multiple types of users will lead to the set value exceeding the limit, resulting in energy waste. Therefore, adapting to the changes in users' indoor temperature needs is the key to achieving precise heating control, improving energy efficiency and reducing emissions.

[0015] In summary, existing heating load forecasting methods either rely on high-cost, sophisticated simulations or black-box learning, making it difficult to balance interpretability, stability, and engineering feasibility under real-world conditions such as limited measurement points, inconsistent measurement standards, variable operating conditions, and sudden weather events. Furthermore, these methods have long weakened the dominant role of room temperature, a core state variable on the user side.

[0016] Therefore, it is necessary to construct a method-oriented heating load forecasting framework. Under the constraint of causal consistency, the framework uses room temperature and its derived comfort characteristics, along with meteorological conditions and their cumulative effects, to form an observable input basis. It should characterize and constrain thermal inertia and uncertain disturbances, and achieve self-learning, self-adaptation, and trend suppression during operation through lightweight rolling correction and near-end incremental updates. It should provide result quality assurance under abnormal and extreme disturbances to ensure that the output does not deteriorate, while supporting time-sharing, zoned, and multi-level deployments. This will improve operating efficiency and reduce energy consumption while meeting users' thermal comfort requirements.

[0017] References [1] Yang Y, Østergaard PA, Wen W, et al. Heating transition in the hot summer and cold winter zone of China: District heating or individual heating?[J]. Energy, 2024, 290: 130283. [2] He Y, Kvan T, Liu M, et al. How green building rating systems affect designing green[J]. Building and Environment, 2018, 133: 19–31. [3] Liu G, Yang H, Fu Y, et al. Cyber-physical system-based real-timemonitoring and visualization of greenhouse gas emissions of prefabricatedconstruction[J]. Journal of Cleaner Production, 2020, 246: 119059. (Nature) [4] Guo Y-Y. Revisiting the building energy consumption in China:Insights from a large-scale national survey[J]. Energy for SustainableDevelopment, 2022, 68: 76–93. [5] Zhao A, Mi L, Xue X, et al. Heating load prediction ofresidential district using hybrid model based on CNN[J]. Energy andBuildings, 2022, 266: 112122. [6] Lin X, Zhang N, Luo Z, et al. Balanced operation strategies ofdistrict heating systems based on dynamic hydraulic-thermal modeling[J].Energy and Built Environment, 2025, 6: 466–483. [7] Hu X, Liu Y, Zhou Y, et al. Prediction and factors determination of district heating load based on random forest algorithm[C] / / The 11thInternational Symposium on Heating, Ventilation and Air Conditioning (ISHVAC2019). Environmental Science and Engineering. Harbin, 2020: 887–895. [8] Zhou Y, Wang L, Qian J. Application of combined models based on onempirical mode decomposition, deep learning, and ARIMA for short-term heatingload predictions[J]. Sustainability, 2022, 14(12): 7349. [9]Reinhart CF, Cerezo Davila C. Urban building energy modeling—Areview of a nascent field[J]. Building and Environment, 2016, 97: 196–202.

[10] Lei Shuyao, Chen Zhenqian. Research on the impact of the external envelope of a residential building on building energy consumption [J]. Building Thermal Energy Ventilation and Air Conditioning, 2020, 39(12):78-82.

[11] Cucca G, Ianakiev A. Assessment and optimization of energy consumption in building communities using an innovative co-simulation tool[J]. Journal of Building Engineering, 2020, 32: 101681.

[12] Khajavi H, Rastgoo A. Improving the prediction of heating energyconsumed at residential buildings using a combination of support vectorregression and meta-heuristic algorithms[J]. Energy, 2023, 272: 127069.

[13] Beijing Energy Heating Co., Ltd. A method for predicting building heating load based on artificial intelligence algorithm: 202510188148.0 [P]. 2025-10-03.

[14] Tianjin University of Technology. A method for short-term load prediction of district heating system based on SD-DNNs: 201910203583.0 [P]. 2023-04-07.

[15] Beijing Heating Group Co., Ltd., Beijing Huare Technology Development Co., Ltd. A method for predicting heating demand load: 201911021471.X[P]. 2024-06-14.

[16] Zhang Lijun, An Wenhan, Zhou Xiaoxuan, et al. Prediction of heating load in a residential area in Kangding City based on improved neural network algorithm [J]. Sichuan Architecture, 2024, 44(06):265-268.

[17] Beijing Jusen Technology Development Co., Ltd. Heating load prediction method and system based on big data: 202410970912.5 [P]. 2025-03-21.

[18] Lu S, Li C, Wang R, Huo Y. Research on the day-ahead scheduling optimization method of medium-depth geothermal cascade heating system[J]. Journal of Building Engineering, 2024, 82: 108184.

[19] Abdou N, El Mghouchi Y, Jraida K, Hamdaoui S, Hajou A, MouqallidM. Prediction and optimization of heating and cooling loads for low energybuildings in Morocco: An application of hybrid machine learning methods[J]. Journal of Building Engineering, 2022, 61: 105332.

[20] Sun C, Liu Y, Cao S, et al. Integrated control strategy of district heating system based on load forecasting and indoor temperature measurement[J]. Energy Reports, 2022, 8: 8124–8139. Summary of the Invention

[0018] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a building cluster heat load prediction method based on adaptive operating conditions and causal residual fusion. It constructs a lightweight, interpretable baseline representation using a small number of observable core variables (indoor temperature, outdoor weather, historical load, and time attributes), and performs minor corrections using data-driven error learning. During operation, key weights and fusion coefficients are incrementally updated via a sliding window, coupled with near-end small-step correction to suppress slow drift. To address the thermal inertia and time-lag effects of the heating system, feature representations with historical memory and cumulative influence are introduced to constrain short-term fluctuations. When the combined result is not better than the baseline within the near-end evaluation window, the correction is automatically discontinued and the baseline output is restored to avoid performance degradation. Thus, under conditions of small sample size and limited measurement points, a continuous and auditable heat load prediction sequence is continuously provided, balancing accuracy, stability, and engineering feasibility.

[0019] The technical problem solved by this invention is achieved through the following technical solution: A method for predicting heat load in building clusters based on adaptive operating conditions and causal residual fusion. The method takes a small amount of observable and traceable operating data as input, and takes into account interpretability, stability and engineering feasibility. It includes data preparation, causal safety feature construction, interpretable baseline modeling, error learning, adaptive synthesis under heating conditions and rolling correction and result quality assurance, forming a rolling prediction sequence oriented to actual operation. The steps of the method are as follows: S1. Data preparation and consistency processing: Align and denoise the raw running data according to the timeline to ensure that only current and historical information is used to meet the causal consistency requirements. S2. Construction of Causal Safety Features: Based on indoor temperature, outdoor temperature and historical load, construct a feature set with few parameters; S3. Interpretable Baseline Modeling: An interpretable regressor is used to fit the features of S2 to obtain the baseline prediction y. base The interpretable regressor includes regularized linear / polynomial regression, piecewise linear / spline regression, generalized linear or additive models; for training tail window on y base Perform monotonic or linear calibration to obtain or its equivalent variants; S4. Error Learner Training: The learning objective is "real load - baseline prediction". A robust regressor is trained as the error channel. The baseline, linearly calibrated at the tail end, is used to capture anomalous perturbations and nonlinear residual modes, and is optimally selected through time-series cross-validation or rolling validation. ; S5. Adaptive Synthesis under Heating Conditions: Within the heating condition subset, the synthesis weight γ of the baseline and error channel is determined based on the prediction error relationship within the recent window; thus, the synthetic prediction model y is obtained. fused ,Right now While maintaining baseline interpretability, it compensates for residual patterns that are difficult to characterize linearly; γ The calculation is performed using least squares, closed-form or numerical solutions with regularization, convex optimization, or Bayesian estimation, and is truncated according to a preset upper limit; during the prediction rolling process, the error learner is periodically retrained and recalculated. γ Furthermore, an online bias correction for the near-end mean is introduced; when the error of the fusion result within the recent window exceeds the allowable proportion of the baseline, a safety backoff is triggered, reducing γ to 0 or decreasing its weight; S6. Rolling Correction and Parameter Update: Near-end mean alignment and periodic window retraining are employed in rolling prediction to suppress slow drift and slight distribution changes, maintaining long-term consistency; including: 1) Proximal mean alignment: at time 1 t During the rolling prediction process, the length of the first proximal window (proximal bias window) is selected as W. b Then at time t The first proximal window is: ; (1) Calculate the mean bias between the baseline calibration output and the actual load within the window: ; (2) where M is the number of samples satisfying data validity within the window; and add the mean bias to the current moment output to complete online bias correction to obtain the corrected load y out :[[]] ; (3) 2) Periodic window retraining and safety fallback: K is the periodic retraining interval, indicating that every K new sampling moments are accumulated during the rolling prediction process to trigger retraining of the residual learner based on the proximal window and update of the fusion coefficient. Let the second proximal window, i.e., the retraining gating window, be W, W ≥ W b , then the second proximal window at time t is:[[]] ; (4) When the update is triggered, retrain the residual learner based on the proximal window Ω(t) and update the fusion coefficient based on the proximal window γ ; and calculate the root mean square error RMSE of the fusion output and the baseline output relative to the true load within the proximal window fused and RMSE base , when the following is satisfied:[[]] ; (5) Trigger safety fallback, and update the fusion coefficient to γ ← 0 or γ ← γ, where: ε ≥ 0 is the allowable proportional parameter, and κ ∈ (0,1) is the attenuation coefficient; 3) Tail segment linear calibration: Perform a linear calibration on y base and y fused respectively at the training tail quantile window to suppress slow drift; S7. Result quality guarantee: When the synthetic prediction is not better than the baseline reference within the recent window, automatically switch to the baseline output or reduce the synthetic weight to ensure that the worst case does not deteriorate under data fluctuations and abnormal disturbances; specifically:[[]] 1) Causal evaluation: At the current moment, only construct a proximal evaluation window based on historical data, and calculate the quality metrics of the baseline prediction and the fusion prediction, including at least one error metric; that is, at the current moment t, only form a window with historical data ∩ "measured set", and count the number of valid samples N; if N < Nmin, then maintain the fusion weight γ of the previous cycle unchanged and end this round; 2) Deterioration determination: When the error of the fusion prediction exceeds the preset tolerance range and deteriorates relative to the baseline performance, it is determined that the fusion enters an unstable state; 3) Gating Execution: Trigger the protection strategy, reduce or freeze the fusion weight γ, or directly switch to the baseline output, and maintain this state during the protection period to avoid frequent jitter, i.e.: Hard shutdown: ; Soft downgrading: ; Simultaneously set a hold counter Where H is the protection step size; 4) Rolling Review: The evaluation window and quality indicators are continuously updated as new data arrives. For each new sample: hold←hold-1, the output remains the same. γ represents the value that is turned off / de-weighted; synchronous scrolling refresh. W t With measurement, periodic correction; for missing values ​​within the window, only one-sided safe filling or removal is performed; 5) Recovery Decision: When the fusion result is better than the baseline again and meets the hysteresis requirement within a continuous window, the protection is lifted and the adaptive update and normal output of γ are restored. 6) Record traceability: Log the window range, quality indicators and gate control actions for evidence storage, providing a basis for subsequent auditing and parameter tuning.

[0020] Furthermore, the specific principle for constructing input features in S2 is as follows: (1) Observability and few parameters: Only core quantities that can be directly measured or recorded in daily operation and maintenance are used, including indoor temperature ( T in ), outdoor temperature ( T out Historical heat load (Q) and time / rhythm information, independent of high-cost information on building geometry and envelope details; (2) Causal consistency and leakage prevention: Features at any given time are constructed solely from current and historical data, and the use of future information is prohibited. All smoothing / difference / memory terms are constructed on one side. (3) Heating specificity: Construct heat demand-related characteristics around the "heating" working condition, and describe thermal inertia and time lag effects with temperature zone, continuity and memory; Under the above principles, the determined input features include: To avoid noise and aperture fluctuations, a one-sided exponential smoothing (causal EMA) is applied to the temperature quantity. t For discrete sampling times, x t This is the original temperature measurement. If the causal smoothing value is given, then: (6) ; in, α For smoothing coefficients,α A larger value indicates a higher weight for the current measurement value; the value range is between 0 and 1. Based on this, a family of features with few parameters is constructed: (a) Temperature difference intensity term (main driver of heat demand), used to characterize the direct impact of indoor-outdoor temperature gradient on instantaneous load; ; (7) Where: t is time. These are the causal smoothing values ​​of the indoor and outdoor temperatures at time t, respectively, which are calculated using formula (6). (b) Heating demand intensity item, heating trigger reference temperature T b The heating degree-day reference temperature can be used to describe the monotonic relationship of "the colder it is, the stronger the heat demand," which is defined as: ; (8) in, For the heating demand intensity term, T b Preset temperature; (c) Characterization of persistence and thermal inertness, Duration L h,t :L h,t = Continuous counting ( T out , t ( s )< T b ) or L h,t = Continuous counting ( T out , t ( s )> T b This reflects the "cumulative increase" in load caused by the duration of heating maintenance; Thermal inertia time delay S t This describes the slow thermal inertia of a period that has been relatively cold / warm recently. ; (9) Where: ρ is the inertia parameter, preferably in the range of 0.85~0.98, T ref The reference temperature is used as the set value, and the user's required temperature is used as the set value. The "heat storage-persistence" difference term is used to express the gradual change in environmental temperature and short-term deviations. ; (10) in, The difference between "long-term background and short-term deviation" reflects the slowly varying warm and cold background and short-term disturbances; Wl To depict the length of the long window and the gradually changing background, W s The short window length characterizes short-term fluctuations; W l and W s Based on empirically given sampling step size and system thermal inertia, and with the optimal selection through time-consistent rolling verification; to balance stability and response speed, W is chosen without increasing the number of measurement points. l ≥4W s ; (d) Temporal and rhythmic features Zt: Temporal and rhythmic features are used to characterize diurnal patterns, intra-weekly differences, holiday disturbances, and long-term slow variations in a low-dimensional manner without increasing the number of measurement points. ; (11) Where: Φ(·) is a periodic basis function mapping, which can take [sinθ,cosθ] or be extended to [sin(nθ),cos(nθ)] of order n≤N for each angle θ. It is recommended that N be minimized to take the value of N=1 or 2. I(t) represents the system / special day instruction (weekends / statutory holidays / make-up workdays, etc., 0 / 1); n t To accumulate the day sequence, count the number of days from the start date to the date at time t; day-level data is generally 1, 2, 3, ...; hour-level data is counted according to "the day number"; Long-term slowly changing bases (trends / slow drifts) can be represented in low dimension using multi-segment linear, low-order polynomial, or spline bases based on cumulative day series. h t This is the intraday slot (hourly data is usually taken as 0~H-1), H=24; if it is daily data, this item can be omitted; w t This is the week index, representing the position number of the date to which time t belongs within a week, with a value ranging from 0 to 6; d t The date sequence within the year indicates that time t is the dth day of the year. t Heaven, d t It is an integer; || is used for concatenating vectors.

[0021] The advantages and beneficial effects of this invention are as follows: (1) Few parameters and easy to obtain: It mainly uses a few core observable measurements such as indoor temperature, outdoor temperature, historical load and time attributes, reducing the reliance on high-dimensional sensing and high-cost modeling information, which makes it easy to be implemented quickly in existing (including old) centralized heating systems.

[0022] (2) Causal consistency and auditability: Features and training - the prediction process strictly follows the time causality to avoid leakage of forward information and data; the output is traceable, easy to operate and maintain for review and supervision.

[0023] (3) Explainable main trend expression: The backbone model provides clear physical meaning and engineering interpretability, forming a robust reference baseline and reducing the decision-making risks brought about by the "black box".

[0024] (4) Explicit characterization of thermal inertia and time delay: The thermal inertia and time delay effect of buildings and systems are reflected by structural quantities such as duration, phased trends and historical memory, so as to improve the continuity and stability under heating conditions.

[0025] (5) Abnormal disturbance compensation capability: Without changing a few parameters, it can learn to compensate for abnormal residuals such as meteorological changes, sudden load changes and measurement fluctuations, thereby improving the prediction accuracy of peak and valley and disturbance scenarios.

[0026] (6) Self-learning and self-adaptation: The windowed small-step update and weight self-determination strategy can automatically correct itself as the running strategy changes and slight caliber drift occurs, reducing manual modeling and frequent retraining.

[0027] (7) Performance lower bound constraint: Set the lower bound constraint of output quality. When the recent performance fails to meet the standard, it will automatically converge to the conservative reference to ensure that the worst performance does not deteriorate under data fluctuations and abnormal disturbances.

[0028] (8) Small sample friendly and quick online: a usable model can be formed with only limited data from the first few days or weeks, shortening the deployment cycle and reducing maintenance costs.

[0029] (9) Strong engineering compatibility: It is suitable for time-sharing and zone-sharing applications of multi-level control units such as heat source-heat exchange station-terminal; it is not sensitive to the degree of system digitization, "it can be used as soon as it can be measured".

[0030] (10) Significant comprehensive benefits: While meeting heating needs and thermal comfort, it reduces the risk of oversupply and peak pressure, supports scheduling optimization and energy-saving assessment, and improves the economy and reliability of the centralized heating system.

[0031] (11) High flexibility: This method can be deployed independently by site or zone, or it can be deployed collaboratively at multiple levels of heat source-heat exchange station-terminal; the rolling cycle can be set at the hourly or daily level; the specific model form of the baseline and error learner, the synthetic weight estimation and window length, etc. can all be equivalently replaced and adjusted without deviating from the above principles. Attached Figure Description

[0032] Figure 1 This is a flowchart of the present invention; Figure 2This is an application flowchart of an embodiment of the present invention; Figure 3 This is a diagram showing the verification set results of an embodiment of the present invention; Figure 4 This is a diagram showing the results of different revision cycles in embodiments of the present invention; Figure 5 This is a graph showing the predicted future heat load under the multi-room temperature target strategy of this invention. Detailed Implementation

[0033] The present invention will be further described in detail below through specific embodiments. The following embodiments are merely descriptive and not limiting, and should not be used to limit the scope of protection of the present invention.

[0034] like Figure 1 , 2 As shown, a building cluster heat load prediction method based on adaptive operating conditions and causal residual fusion is innovative in that: this embodiment selects the operating data of an energy station in Tianjin during two consecutive heating seasons, Heating Season 1 and Heating Season 2, as samples. Heating Season 1 is used for model building as the training set, while Heating Season 2 serves as an independent validation set to examine the prediction of future loads. Input data includes: daily heat load Q. real Indoor temperature T in Outdoor temperature T out With available calendar information.

[0035] The specific implementation steps are as follows: S1. Data Preparation and Causal Consistency Processing: Sort and align the original data by timestamp, delete significant anomalies and unusable records, and ensure that feature construction and training only use current and historical information to satisfy causal consistency.

[0036] S2, the characteristic structure of causal safety: temperature smoothness and temperature difference: for T in , T out T is obtained using the causal exponential moving average (EMA). in,s ,T out,s ; make ΔT=T in,s -T out,s It represents the intensity of immediate demand.

[0037] Heating intensity and duration: Heating intensity is defined as H = max(0, T) base -T out,s ), T base =18℃ and count the number of times the temperature is continuously below the reference temperature as the duration, which is used to reflect the cumulative effect of thermal inertia / time lag.

[0038] Continuous-gradually varying composite quantity: the difference in smoothness of outdoor temperature difference between long / short windows. ; In this embodiment, Ws=3 is calculated, W l =21 In this example ρ =0.95, T ref =20℃; Time and rhythm: If timestamps are available, extract the sine and cosine bases of low-order intraday / weekly / seasonal phases (and weekend / holiday indicators) to express rhythms and institutional differences.

[0039] Historical Memory: Incorporating First-Order Load Delay Q t-1 To reflect the short-term regression effect.

[0040] All of the above features are low-dimensional, causal, observable, and interpretable, avoiding reliance on high-cost modeling information.

[0041] S3. Explainable baseline modeling (main trend) uses a lightweight interpretable pipeline of multinomial features (second order) – standardized – ridge regression to fit the features of S2, obtaining the baseline prediction y. base To improve the fit to the statistical bias at the end of the training set, a light linear calibration is performed at the end of the training set: =a0 y base +b0.

[0042] S4. Error channel training aims to achieve "real load - baseline prediction (after calibration)". Based on time series cross-validation, it selects the best among several robust regressors (ridge regression, Huber regression, elastic net) to obtain residual predictions. r ^To compensate for anomalous disturbances and nonlinear residual modes.

[0043] S5, Adaptive Synthetic Model under Heating Conditions In the heating condition subset, based on the recent window (y) true -y base_cal )and Relationship, estimate the composite weights γ (Constrain the upper limit to prevent overfitting); obtain the synthetic prediction. ; The entire process is adaptive on the heating subset.

[0044] S6. Rolling Prediction and Parameter Update (Small-Step Adaptive) In rolling applications: (i) mean alignment of near-end small windows is used to suppress mild caliber drift; (ii) windowed small window retraining and weight reestimation are performed with a fixed step size to absorb slowly changing operating conditions; (iii) residual delay features are updated to maintain causal closure.

[0045] S7. Lower limit constraint on result quality If the error of the synthesized result relative to the baseline does not improve (or deteriorates beyond the threshold) within the recent window, the weighting will be automatically reduced / switched back to the baseline output to ensure that the prediction performance does not deteriorate under data fluctuations and abnormal disturbances.

[0046] S8, Output and Archive Export rolling prediction sequences, training / validation metric reports, and comparison charts (real-baseline-synthetic), and write them to disk in time sequence for operation and maintenance auditing and review.

[0047] This embodiment selects energy station operation data of a centralized heating system during the heating season as a sample. The input only relies on a small number of observable measurements such as indoor temperature, outdoor temperature, historical heat load, and time attributes. The data is aligned by timestamp and denoised to strictly maintain causal consistency. For the independent validation set (heating season 2), all model parameters obtained during the training phase are retained without any retraining or parameter tuning.

[0048] As shown in Table 1, the validation results on the validation set show that the baseline model achieves RMSE=12.221, MAE=9.615, and R0. 2 =0.8842, MAPE=10.38%; based on this, the output of this dynamic prediction model reaches RMSE=9.488, MAE=7.445, R 2 =0.9422, MAPE=7.87%. Compared to the baseline, the synthesized results achieved a reduction of approximately 22.4% in RMSE, approximately 22.5% in MAE, and approximately 24.2% in MAPE on this independent validation set. 2 An improvement of 0.038. The above results were obtained under conditions relying only on a limited number of observables such as indoor temperature, outdoor temperature, historical load, and time attributes. This demonstrates the improved accuracy and engineering verifiability of the method in small sample and real-world data scenarios. The results are shown in Table 1 and... Figure 3 As shown.

[0049] Table 1 Evaluation criteria results for different models

[0050] In this embodiment, the outdoor temperature sequence of heating season 2 is used as the future meteorological input, and the room temperature targets are set at 18℃, 20℃, 22℃, 24℃, and 26℃, respectively, thereby obtaining multiple corresponding load change trend curves. At the same time, in order to more closely reflect the actual operation scenario and reflect the periodic correction capability of the method after the arrival of new heating data, the heating season 2 data is continuously input into the model for online revision, and future load prediction is carried out on this basis to form a multi-scenario prediction set, providing a unified data foundation for subsequent target energy consumption management, zoned room temperature scheduling, and strategy comparison.

[0051] From Table 2 and Figure 4 It can be seen that when the revision period increases from 3 days to 14 days, the RMSE decreases from 9.65 MWh to 8.80 MWh, and R... 2 The increase from 0.940 to 0.951 indicates that appropriately extending the revision window helps suppress the interference of short-term noise on bias estimation and improves the stability of dynamic revision; however, when the period continues to increase to 20 days and 30 days, the RMSE rises back to 8.95 MWh and 9.09 MWh, respectively. 2 The slight decrease indicates that an excessively long window weakens the model's response speed to changes in load structure and operating conditions. Therefore, the effectiveness of dynamic model revision is not simply a matter of "the shorter the interval, the better." Instead, it requires effective identification and balancing of deviations / drifts based on load change characteristics. Under these data conditions, a 14-day revision cycle achieved the best overall performance and can be considered one of the recommended ranges for engineering deployment in this scenario.

[0052] Table 2. Model evaluation criteria results under different periodicity correction windows.

[0053] Figure 5 This presents the future heat load forecast results under a multi-room-target strategy. Based on the optimal revision window, 28 days of data from heating season 2 are input for model revision, and a further 3-week (21-day) heat load forecast is then performed. Figure 5 As shown, the horizontal axis represents the sample number (corresponding to day / time period), and the vertical axis represents the heat load. In the historical segment, a comparison between the actual heat load and the model's simulated values ​​is presented to demonstrate the model's ability to fit load changes under known operating conditions. In the future segment, using the outdoor temperature sequence of heating season 2 as future meteorological input, indoor temperature targets of 18℃, 20℃, 22℃, 24℃, and 26℃ are set under the same outdoor temperature conditions, resulting in corresponding multi-scenario load prediction curves.

[0054] As can be seen from the future section, the higher the room temperature target, the higher the overall predicted load level, and the curves of each scenario maintain a stable size pattern (highest at Tin=26℃, lowest at Tin=18℃). This indicates that the method can directly output the load trend and difference magnitude corresponding to different room temperature strategies under the same meteorological boundary. At the same time, the scenario curves show synchronous changes during the load climbing and fluctuation stages, indicating that the model has a consistent response to the main load trend driven by outdoor temperature, while the room temperature target mainly reflects the systematic increase in load level. This provides a set of multi-scenario predictions that can be directly called upon for subsequent target energy consumption control, regional room temperature scheduling, and strategy comparison.

[0055] During actual operation and adjustment, based on the time-varying and variability of users' actual needs and future meteorological factors, the system can output rolling heat load forecasts and target load sequences for scheduling and energy consumption management from operation and environmental observation data (such as outdoor weather, indoor thermal environment, equipment operation records, historical loads, time series information, etc.). It can also be accompanied by confidence, alarm and diagnostic information to support operation decisions and performance accounting. It supports individual or layered implementation at any level of the building from heat source to heat exchange station to terminal. It can be connected to mature information / digital transformation systems, or connected to traditional systems with complete basic measurement points with minimal modifications, covering rolling applications from hourly to daily levels. As long as the input parameters meet the time scale uniformity, modeling can be achieved. An executable target load is formed with the target energy consumption as the driving force. It has interpretability and traceability, lightweight and rapid deployment, insensitivity to the degree of digitalization (measurable and usable) and cross-scenario transferability, thereby shortening the online cycle and reducing maintenance costs.

[0056] Typical applicable scenarios include time-sharing and zone-based load, stable prediction under different operating states (initial heating - steady state - warming transition stage, extreme cold wave and operating condition switching) and robust prediction and target generation under uncertain system disturbances, as well as heating performance evaluation, energy saving benchmarking and target energy consumption closed loop.

[0057] Although embodiments and drawings of the present invention have been disclosed for illustrative purposes, those skilled in the art will understand that various substitutions, variations and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. A method for predicting the heat load of building complexes based on adaptive operating conditions and causal residual fusion, characterized in that: The method takes a small, observable, and traceable amount of operation as input, and takes into account interpretability, stability, and engineering feasibility. It includes method data preparation, causal and safe feature construction, interpretable baseline modeling, error learning, adaptive synthesis under heating conditions, rolling correction, and result quality assurance, forming a rolling prediction sequence oriented towards actual operation. The steps of the method are as follows: S1. Data preparation and consistency processing: Align and denoise the raw running data according to the timeline to ensure that only current and historical information is used to meet the causal consistency requirements. S2. Construction of Causal Safety Features: Based on indoor temperature, outdoor temperature and historical load, construct a feature set with few parameters; S3. Interpretable Baseline Modeling: An interpretable regressor is used to fit the features of S2 to obtain the baseline prediction y. base The interpretable regressor includes regularized linear / polynomial regression, piecewise linear / spline regression, generalized linear or additive models; for training tail window on y base Perform monotonic or linear calibration to obtain or its equivalent variants; S4. Error Learner Training: The learning objective is "real load - baseline prediction". A robust regressor is trained as the error channel. The baseline, linearly calibrated at the tail end, is used to capture anomalous perturbations and nonlinear residual modes, and is optimally selected through time-series cross-validation or rolling validation. ; S5. Adaptive Synthesis under Heating Conditions: Within the heating condition subset, the synthesis weight γ of the baseline and error channel is determined based on the prediction error relationship within the recent window; thus, the synthetic prediction model y is obtained. fused ,Right now While maintaining baseline interpretability, it compensates for residual patterns that are difficult to characterize linearly; γ The calculation is performed by least squares, closed or numerical solutions with regularization, convex optimization or Bayesian estimation, and truncated by a preset upper limit. During the prediction rolling process, the error learner is periodically retrained and recalculated. γ Furthermore, an online bias correction for the near-end mean is introduced; when the error of the fusion result within the recent window exceeds the allowable proportion of the baseline, a safety backoff is triggered, reducing γ to 0 or decreasing its weight; S6. Rolling Correction and Parameter Update: Near-end mean alignment and periodic window retraining are employed in rolling prediction to suppress slow drift and slight distribution changes, maintaining long-term consistency; including: 1) Proximal mean alignment: at time 1 t During the rolling prediction process, the length of the first proximal window (proximal bias window) is selected as W. b Then at time t The first proximal window is: ； （1） Calculate the mean bias between the baseline calibration output and the actual load within the window: ； （2） Where M is the number of samples within the window that meet the data validity requirements; and the mean bias is added to the output at the current time to complete the online bias correction and obtain the corrected load y. out : ； （3） 2) Periodic window retraining and safety rollback: K is the periodic retraining interval, which means that during the rolling prediction process, a retraining of the residual learner based on the near-end window and an update of the fusion coefficients are triggered every K new sampling times. Let the second near-end window, i.e. the retraining gate window, be W, where W ≥ W b Then at time t The second proximal window is: ； （4） Upon triggering an update, the residual learner is retrained based on the proximal window Ω(t), and the fusion coefficients are updated based on the proximal window. γ Within the near-end window, the root mean square error (RMSE) of the fused output and the baseline output relative to the actual load is calculated respectively. fused With RMSE base When the following conditions are met: ； （5） Trigger a safety rollback and update the fusion coefficient to γ←0 or γ← γ, where: ε≥0 is the allowable scaling parameter, and κ∈(0,1) is the attenuation coefficient; 3) Tail-end linear calibration: Within the training tail-end quantile window, y... base With y fused Perform a linear calibration once each To suppress slow drift; S7. Result Quality Assurance: When the composite prediction is not better than the baseline reference within the recent window, automatically switch to the baseline output or reduce the composite weight to ensure that the worst-case scenario does not deteriorate under data fluctuations and abnormal disturbances; specifically: 1) Causal assessment: At the current moment, only construct a proximal assessment window based on historical data, calculate the quality metrics of the baseline prediction and the fusion prediction respectively, including at least one error metric; that is, at the current moment t, only form a window with historical data ∩ "measured set", and count the number of valid samples N; if N < Nmin, then keep the fusion weight γ of the previous cycle unchanged and end this round; 2) Degradation judgment: When the error of the fusion prediction exceeds the preset tolerance range and the performance degrades relative to the baseline, the fusion is judged to have entered an unstable state; 3) Gating Execution: Trigger the protection strategy, reduce or freeze the fusion weight γ, or directly switch to the baseline output, and maintain this state during the protection period to avoid frequent jitter, i.e.: Hard shutdown: ; Soft downgrading: ; Simultaneously set a hold counter Where H is the protection step size; 4) Rolling Review: The evaluation window and quality indicators are continuously updated as new data arrives. For each new sample: hold←hold-1, the output remains the same. γ represents the value that is turned off / de-weighted; synchronous scrolling refresh. W t With measurement, periodic correction; for missing values ​​within the window, only one-sided safe filling or removal is performed; 5) Recovery Decision: When the fusion result is better than the baseline again and meets the hysteresis requirement within a continuous window, the protection is lifted and the adaptive update and normal output of γ are restored. 6) Record traceability: Log the window range, quality indicators and gate control actions for evidence storage, providing a basis for subsequent auditing and parameter tuning.

2. The building cluster heat load prediction method based on adaptive operating conditions and causal residual fusion as described in claim 1, characterized in that: The specific principle for constructing input features in S2 is as follows: (1) Observability and few parameters: Only core quantities that can be directly measured or recorded in daily operation and maintenance are used, including indoor temperature. T in Outdoor temperature T out Historical heat load Q and time / rhythm information, independent of high-cost information on building geometry and envelope details; (2) Causal consistency and leakage prevention: Features at any given time are constructed solely from current and historical data, and the use of future information is prohibited. All smoothing / difference / memory terms are constructed on one side. (3) Heating specificity: Construct heat demand-related characteristics around the "heating" working condition, and describe thermal inertia and time lag effects with temperature zone, duration and memory; Under the above principles, the determined input features include: To avoid noise and aperture fluctuations, a one-sided exponential smoothing (causal EMA) is applied to the temperature quantity. t For discrete sampling times, x t This is the original temperature measurement. If the causal smoothing value is given, then: （6） ； in, α For smoothing coefficients, α A larger value indicates a higher weight for the current measurement value; the value range is between 0 and 1. Based on this, a family of features with few parameters is constructed: (a) Temperature difference intensity term / main driver of heat demand, used to characterize the direct impact of indoor-outdoor temperature gradient on instantaneous load; ； （7） Where: t is time. These are the causal smoothing values ​​of the indoor and outdoor temperatures at time t, respectively, and are specifically calculated using formula (6); (b) Heating demand intensity item, heating trigger reference temperature T b The heating degree-day reference temperature can be used to describe the monotonic relationship of "the colder it is, the stronger the heat demand," which is defined as: ； （8） in, For the heating demand intensity term, T b Preset temperature; (c) Characterization of persistence and thermal inertness, Duration L h,t :L h,t = Continuous counting ( T out , t ( s )< T b ) or L h,t = Continuous counting ( T out , t ( s )> T b This reflects the "cumulative increase" in load caused by the duration of heating. Thermal inertia time delay S t This describes the slow thermal inertia of a period that has been relatively cold / warm recently. ； （9） Where: ρ is the inertia parameter, preferably ranging from 0.85 to 0.98, T ref The reference temperature is used as the set value, and the user's required temperature is used as the set value. The "heat storage-persistence" difference term is used to express the gradual change in environmental temperature and short-term deviations. ； （10） in, The difference between "long-term background and short-term deviation" reflects the slowly varying warm and cold background and short-term disturbances; W l To depict the length of the long window and the gradually changing background, W s The short window length characterizes short-term fluctuations; W l and W s Based on empirically given sampling step size and system thermal inertia, and with the optimal selection through time-consistent rolling verification; to balance stability and response speed, W is chosen without increasing the number of measurement points. l ≥4W s ; (d) Temporal and rhythmic features Zt: Temporal and rhythmic features are used to characterize diurnal patterns, intra-weekly differences, holiday disturbances, and long-term slow variations in a low-dimensional manner without increasing the number of measurement points. ； （11） Where: Φ(·) is a periodic basis function mapping, which can take [sinθ,cosθ] or be extended to [sin(nθ),cos(nθ)] of order n≤N for each angle θ. It is recommended that N be minimized to take the value of N=1 or 2. I(t) represents the system / special day instruction (weekends / statutory holidays / make-up workdays, etc., 0 / 1); n t To accumulate the day sequence, count the number of days from the start date to the date at time t; day-level data is generally 1, 2, 3, ...; hour-level data is counted according to "the day number"; Long-term slowly changing bases (trends / slow drifts) can be represented in low dimension using multi-segment linear, low-order polynomial, or spline bases based on cumulative day series. h t This is the intraday slot (hourly data is usually taken as 0~H-1), H=24; if it is daily data, this item can be omitted; w t This is the week index, representing the position number of the date to which time t belongs within a week, with a value ranging from 0 to 6; d t The date sequence within the year indicates that time t is the dth day of the year. t Heaven, d t It is an integer; || is used for concatenating vectors.

Images