Intelligent temperature control method and system based on thermal load prediction

Abstract

The invention relates to the field of intelligent temperature control, and particularly discloses an intelligent temperature control method and system based on thermal load prediction, and the method comprises the steps: firstly carrying out the modal exclusive extraction and cross-modal feature fusion of a time sequence and image data flow, and constructing a multi-modal representation to synchronously output a predicted thermal load and uncertainty quantification result; on the basis, a dynamic risk factor is calculated by coupling a current thermodynamic buffer space and predicting uncertainty, and then the composite weight of feedforward and feedback control is adjusted in a self-adaptive mode. According to the mechanism, the energy-saving advantage of feed-forward adjustment is maximized while the safety bottom line is guaranteed, and the technical problems that single-mode prediction is inaccurate and the robustness of a static control strategy is poor are effectively solved.

Description

Technical Field

[0001] This application relates to the field of intelligent temperature control, and more specifically, to an intelligent temperature control method and system based on heat load prediction. Background Technology

[0002] With the rapid development of modern building automation technology, achieving efficient energy utilization while ensuring the thermal comfort of indoor occupants has become a core requirement in the field of building control. Traditional temperature control systems often employ passive response control logic (such as PID control) or rigid strategies based on fixed schedules, typically only adjusting after the indoor temperature has deviated from the set value. This lag not only leads to energy waste but also makes it difficult to cope with complex and changing human activities and environmental disturbances. Therefore, constructing an intelligent temperature control solution based on heat load prediction, which proactively anticipates future heat demands and makes pre-adjustments, has become a key technological path to solving the above problems.

[0003] However, existing temperature control solutions based on heat load forecasting still have significant limitations when dealing with complex application scenarios. Although some technologies attempt to introduce machine learning models for load forecasting, they often struggle to effectively integrate multimodal heterogeneous data. In real-world building environments, factors influencing heat load include not only traditional temperature and humidity time-series data but also unstructured data such as crowd density images from surveillance cameras. Existing technologies often rely on manual feature engineering to forcibly compress these heterogeneous data into flat numerical features, resulting in the significant loss of spatial distribution information and temporal semantic information in the images, creating a semantic gap. This lack of a mechanism for deep cross-modal interactive learning makes it difficult for models to capture the nonlinear synergistic effects between multi-source information, thus limiting the immediacy and accuracy of load forecasting.

[0004] More importantly, traditional predictive models typically only output deterministic point predictions, lacking an assessment of the reliability of the prediction results. In actual control, blindly trusting an uncertain prediction can lead to erroneous control decisions. Furthermore, even with the uncertainty of the prediction, existing control strategies often fail to dynamically adjust based on the system's current actual safety status. The numerical value of prediction uncertainty alone cannot fully reflect control risk, because the same prediction deviation may be insignificant when the system is in the comfort zone, but could lead to serious overshooting if the system is at the edge of the comfort zone. Existing technologies lack a mechanism to couple prediction variance with the current thermodynamic buffer space, making it impossible to dynamically balance the aggressiveness of feedforward control with the conservatism of feedback control based on real-time risk, and thus difficult to maximize the energy-saving benefits of predictive control while ensuring a safety baseline.

[0005] Therefore, an optimized intelligent temperature control scheme based on heat load prediction is desired. Summary of the Invention

[0006] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide an intelligent temperature control method and system based on heat load prediction.

[0007] According to one aspect of this application, a smart temperature control method based on heat load prediction is provided, comprising: Acquire time-series data streams and image data streams; Modality-specific feature extraction is performed on time-series data streams and image data streams to obtain time-series sequence embeddings and visual embeddings; Feature fusion is performed on temporal sequence embeddings and visual embeddings to obtain a multimodal fusion representation vector; Heat load prediction and uncertainty quantification are performed on the multimodal fusion representation vector to obtain the predicted heat load and prediction uncertainty; Based on the uncertainty of prediction, a prediction-based feedforward-feedback coordinated control is performed on the predicted heat load and the current state to obtain actuator commands.

[0008] According to another aspect of this application, an intelligent temperature control system based on heat load prediction is provided, comprising: The real-time data acquisition module is used to acquire time-series data streams and image data streams; The modality-specific feature extraction module is used to extract modality-specific features from time-series data streams and image data streams to obtain time-series sequence embeddings and visual embeddings; The multimodal feature fusion module is used to fuse features from temporal sequence embeddings and visual embeddings to obtain a multimodal fusion representation vector; The heat load and uncertainty prediction module is used to perform heat load prediction and uncertainty quantification on the multimodal fusion representation vector to obtain the predicted heat load and predicted uncertainty. The feedforward-feedback coordinated control module is used to perform prediction-based feedforward-feedback coordinated control on the predicted heat load and current state based on prediction uncertainty to obtain actuator commands.

[0009] Compared with existing technologies, this application provides an intelligent temperature control method and system based on heat load prediction. First, it performs modal-specific extraction and cross-modal feature fusion on time-series and image data streams to construct a multimodal representation, simultaneously outputting the predicted heat load and uncertainty quantification results. Based on this, by coupling the current thermodynamic buffer space with the prediction uncertainty, a dynamic risk factor is calculated, and then the combined weights of feedforward and feedback control are adaptively adjusted. This mechanism maximizes the energy-saving advantages of feedforward regulation while ensuring a safety baseline, effectively solving the technical problems of inaccurate single-modal prediction and poor robustness of static control strategies. Attached Figure Description

[0010] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0011] Figure 1 This is a flowchart of an intelligent temperature control method based on heat load prediction according to an embodiment of this application; Figure 2 This is a schematic diagram of the data flow of the intelligent temperature control method based on heat load prediction according to an embodiment of this application; Figure 3 The flowchart illustrates the process of performing heat load prediction and uncertainty quantification on a multimodal fusion representation vector to obtain the predicted heat load and prediction uncertainty according to the intelligent temperature control method based on heat load prediction in the embodiments of this application. Figure 4 This is a flowchart illustrating the process of obtaining actuator commands by performing prediction-based feedforward-feedback coordinated control on the predicted heat load and the current state based on the prediction uncertainty of the intelligent temperature control method based on heat load prediction according to the embodiments of this application. Figure 5 This is a block diagram of an intelligent temperature control system based on heat load prediction according to an embodiment of this application. Detailed Implementation

[0012] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0013] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0014] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0015] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0016] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0017] To address the technical problems mentioned in the background, this solution proposes an intelligent temperature control method based on heat load prediction. This method first utilizes a hierarchical cross-modal fusion Transformer architecture to process multi-source heterogeneous data. Specifically, independent intra-modal encoders extract the dynamic features of time-series data and the spatial features of image data, eliminating inter-modal heterogeneity. Subsequently, a cross-modal attention mechanism is used to deeply mine the nonlinear synergistic effect between visual and time-series information (such as the lag effect of personnel density images on heat load changes), outputting a multi-modal fusion representation containing the predicted mean and variance, thereby simultaneously achieving accurate heat load prediction and reliable quantification. Building upon this, an adaptive control strategy based on state-dependent risk assessment is further introduced. This strategy does not simply rely on the predicted uncertainty value, but couples it with the current system's thermodynamic buffer space (i.e., the safety margin between the current temperature and the comfort zone boundary) to calculate a dynamic risk factor. This risk factor adaptively adjusts the combined weights of feedforward and feedback control: when the system is in a safe zone and predictions are reliable, it favors aggressive feedforward control to maximize energy savings; when the system approaches the edge of the comfort zone or predictions are questionable, it automatically and smoothly switches to conservative feedback control to ensure a safety baseline. This mechanism effectively solves the technical challenges of inaccurate single-modal predictions and poor robustness of static control strategies under critical conditions. The technical solution of this application proposes an intelligent temperature control method based on heat load prediction. Figure 1 This is a flowchart of an intelligent temperature control method based on heat load prediction according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow of an intelligent temperature control method based on heat load prediction according to an embodiment of this application. Figure 1 and Figure 2As shown, the intelligent temperature control method based on heat load prediction according to an embodiment of this application includes the following steps: S100, acquiring a time-series data stream and an image data stream; S200, performing modality-specific feature extraction on the time-series data stream and the image data stream to obtain a time-series sequence embedding and a visual embedding; S300, performing feature fusion on the time-series sequence embedding and the visual embedding to obtain a multi-modal fusion representation vector; S400, performing heat load prediction and uncertainty quantization on the multi-modal fusion representation vector to obtain a predicted heat load and a predicted uncertainty; S500, based on the predicted uncertainty, performing prediction-based feedforward-feedback collaborative control on the predicted heat load and the current state to obtain actuator commands.

[0018] Specifically, in step S100, time-series data streams and image data streams are acquired. It should be understood that single-mode time-series temperature and humidity data can only reflect the delayed results of thermodynamic state changes and cannot intuitively perceive physical triggers that cause rapid changes in heat load, such as sudden gatherings of people, the start of high-density meetings, or direct sunlight. This lack of information in a single mode leads to significant perception blind spots and response delays in traditional prediction models when facing complex dynamic scenarios. Therefore, in the technical solution of this application, time-series data streams and image data streams are further acquired to construct a multi-dimensional perception input space covering basic environmental physical parameters and visual scene semantic features. This allows for the simultaneous capture of numerical information reflecting the current system's thermodynamic response and visual information reflecting potential sources of thermal disturbance. This provides a comprehensive, heterogeneous, and complementary data foundation for subsequent mode-specific feature extraction and multi-modal fusion, enabling the control system not only to accurately know the current indoor environmental state but also to anticipate future heat load fluctuations through image data, thereby significantly improving the system's predictive ability and control robustness against dynamic heat load changes.

[0019] More specifically, in a particular example of this application, the intelligent temperature control system establishes communication connections with environmental sensing devices and visual acquisition devices distributed within the target temperature control area through a building automation network and a security monitoring network, respectively, and performs parallel data acquisition operations. For acquiring time-series data streams, the system uses the BACnet industrial communication protocol to directly read real-time readings from the building management system of indoor temperature sensors, indoor humidity sensors, and carbon dioxide concentration sensors within the target area at a preset fixed sampling frequency, and strictly arranges these continuous numerical data according to timestamp order to organize them into a time-series data stream. For acquiring image data streams, the system uses the RTSP streaming media transmission protocol to capture high-definition video keyframes in real time from monitoring cameras installed on the ceiling or corners of the target area, and organizes these two-dimensional pixel matrix sequences, containing information on the density of people in the area, their activity status, and the opening and closing of curtains, into an image data stream.

[0020] Specifically, in step S200, modality-specific feature extraction is performed on the time-series data stream and image data stream to obtain time-series sequence embeddings and visual embeddings. It should be understood that, due to the significant heterogeneity between the original time-series numerical data and image pixel data in terms of data dimension, sampling frequency, and physical meaning, directly splicing or fusing such low-level heterogeneous data will lead to a severe misalignment of the feature space distribution, making it impossible for the model to effectively capture nonlinear correlations between modalities. Therefore, in the technical solution of this application, modality-specific feature extraction is further performed on the time-series data stream and image data stream to obtain time-series sequence embeddings and visual embeddings. This maps the physically dissimilar original data into the same high-dimensional abstract feature space, thereby eliminating dimensional differences between modalities and extracting feature representations with high-level semantic information. This provides a mathematically computable and semantically aligned standardized input foundation for subsequent multimodal feature fusion, significantly improving the model's comprehensive representation ability of complex environmental states.

[0021] More specifically, in this embodiment, modality-specific feature extraction is performed on the time-series data stream and image data stream to obtain time-series sequence embedding and visual embedding, including: inputting the time-series data stream into a time-series encoder to obtain time-series sequence embedding; and inputting the image data stream into an image feature encoder to obtain visual embedding. It is worth noting that the time-series encoder is a one-dimensional convolutional neural network model, and the image feature encoder is a pre-trained two-dimensional convolutional neural network model. Specifically, the processor of the intelligent temperature control system executes a parallel dual-channel feature encoding program to complete the mapping transformation from raw sensor data to high-dimensional feature vectors. For the processing of the time-series data stream, the system inputs a time-series data stream containing historical continuous numerical sequences of indoor temperature, humidity, and carbon dioxide concentration as the one-dimensional convolutional neural network model of the time-series encoder. This model extracts local trend features and periodic fluctuation patterns by performing sliding convolution operations along the time axis, and outputs a time-series sequence embedding reflecting the dynamic changes of thermodynamic parameters. For image data stream processing, the system inputs the image data stream containing the pixel matrix sequence of monitoring key frames as a pre-trained two-dimensional convolutional neural network model for image feature encoder. This model extracts the texture, shape and spatial distribution semantics in the image layer by layer through multi-layer two-dimensional spatial convolution and pooling operations, and outputs a visual embedding that reflects the density and activity status of people in the area.

[0022] Specifically, in step S300, feature fusion is performed on the temporal sequence embedding and the visual embedding to obtain a multimodal fusion representation vector. It should be understood that, due to the complex cross-modal synergistic relationship between the thermodynamic hysteresis effect in time-series data and the instantaneous state changes in image data, simple linear concatenation cannot effectively capture this deep nonlinear causal chain, making it difficult for the model to understand the specific driving effect of visual scene changes on future heat load. Therefore, in the technical solution of this application, intramodal encoding is further performed on the temporal sequence embedding and the visual embedding to obtain a temporal intramodal correlation encoding vector and a visual intramodal correlation feature vector, and cross-modal encoding is performed on the temporal intramodal correlation encoding vector and the visual intramodal correlation feature vector to obtain a multimodal fusion representation vector. This constructs a unified representation space that first independently refines intramodal features and then deeply interacts with cross-modal information, ensuring that the model can adaptively learn the dynamic weights of visual information for heat load prediction. This significantly enhances the system's ability to understand the semantics of multi-source heterogeneous data in complex scenarios, thereby generating a fusion feature representation that contains rich cross-modal interaction information and is highly sensitive to heat load prediction tasks.

[0023] More specifically, in the embodiments of this application, feature fusion of temporal sequence embedding and visual embedding to obtain a multimodal fusion representation vector includes: intramodal encoding of temporal sequence embedding and visual embedding to obtain a temporal intramodal association encoding vector and a visual intramodal association feature vector; and cross-modal encoding of the temporal intramodal association encoding vector and the visual intramodal association feature vector to obtain a multimodal fusion representation vector.

[0024] In a specific example of this application, the processor of the intelligent temperature control system performs feature fusion tasks through a hierarchical, cascaded attention mechanism architecture. In the first stage of intramodal encoding, the system embeds the temporal sequence generated in the preceding steps into an independent temporal Transformer encoder. This encoder utilizes a multi-head self-attention mechanism to calculate the temporal dependencies within the temporal data, outputting a temporal intramodal association encoding vector that enhances the dynamic features of the temporal dimension. Simultaneously, the system inputs the visual embedding into an independent visual Transformer encoder in parallel. This encoder utilizes a spatial self-attention mechanism to calculate the spatial positional dependencies between image patches, outputting a visual intramodal association feature vector that enhances spatial semantic features. In the second stage of cross-modal encoding, the system concatenates the two refined vector sequences and adds modality type embeddings and position embeddings to construct a hybrid sequence, which is then input into the cross-modal Transformer encoder. This cross-modal encoder calculates the mutual attention weights between temporal features and visual features through a global self-attention mechanism, thereby capturing the nonlinear correlation between image features such as a surge in population density and subsequent temporal features of rising indoor temperature, and finally extracts the vector corresponding to the classification label in the mixed sequence as a multimodal fusion representation vector.

[0025] Specifically, in step S400, the multimodal fusion representation vector is subjected to heat load prediction and uncertainty quantification to obtain the predicted heat load and prediction uncertainty. It should be understood that since the temperature control environment is essentially a nondeterministic system full of random disturbances, a single point prediction value (such as simply predicting a heat load of 1.5kW) cannot reflect the model's confidence level in the current prediction result. Blindly accepting a high-risk prediction value may lead to the collapse of the robustness of the control strategy. Therefore, in the technical solution of this application, the multimodal fusion representation vector is further subjected to heat load prediction and uncertainty quantification to obtain the predicted heat load and prediction uncertainty. This constructs a prediction output mechanism based on probability distribution, which not only provides an expected estimate of the future heat load but also simultaneously quantifies the confidence range of this estimate. This provides a key basis for downstream control systems to assess prediction risks, enabling them to distinguish between deterministic predictions and speculative predictions, thereby achieving safer and more robust decision-making in the face of complex and ever-changing dynamic environments.

[0026] Figure 3 This document describes a flowchart illustrating the process of predicting heat load and uncertainty in a multimodal fusion representation vector using a smart temperature control method based on heat load prediction, according to embodiments of this application. Figure 3As shown, step S400 includes: S410, inputting the multimodal fusion representation vector into the multilayer perceptron used for regression prediction to obtain the original output vector; S420, performing probability distribution parameter analysis and constraint on the original output vector to obtain the predicted heat load and prediction uncertainty.

[0027] Accordingly, in step S410, the multimodal fusion representation vector is input into the multilayer perceptron used for regression prediction to obtain the original output vector. It should be understood that although the multimodal fusion representation vector is rich in semantic information of the scene, it is essentially still an abstract high-dimensional feature representation, which cannot directly correspond to specific thermodynamic values ​​and uncertainty measures in the physical world. Furthermore, there is an extremely complex nonlinear mapping relationship between this high-dimensional feature and the final probability distribution parameters. Therefore, in the technical solution of this application, the multimodal fusion representation vector is further input into the multilayer perceptron used for regression prediction to obtain the original output vector, thereby constructing a dedicated feature decoding and parameter regression channel. The general function approximation capability of the multilayer perceptron is used to accurately transform the abstract semantic features into low-level parameters that control the shape of the probability distribution. This provides the necessary numerical foundation for subsequent probability distribution construction, thus achieving the crucial leap from semantic understanding to physical prediction.

[0028] Specifically, in one example of this application, the processor of the intelligent temperature control system executes a mapping and decoding procedure from high-dimensional features to low-dimensional parameters. First, the system inputs a 768-dimensional multimodal fusion representation vector, output from an upstream cross-modal Transformer encoder and containing temporal and visual depth interaction information, into a specially constructed multilayer perceptron for regression prediction. This multilayer perceptron consists of an input layer, two hidden layers containing ReLU activation functions, and a cascaded linear output layer. During the forward propagation of the data, the multimodal fusion representation vector undergoes linear weighted summation and nonlinear transformation layer by layer in the hidden layers, compressing and reconstructing the original high-dimensional feature space into a more compact latent semantic space. Finally, after a fully connected operation in the linear output layer, the system outputs a two-dimensional real-valued vector without any activation function constraints, i.e., the original output vector, where the first dimension of this vector corresponds to the mean precursor of the heat load prediction, and the second dimension corresponds to the logarithmic precursor of the variance of the heat load prediction.

[0029] Accordingly, in step S420, the original output vector is analyzed and constrained for probability distribution parameters to obtain the predicted heat load and prediction uncertainty. It should be understood that since the original numerical values ​​output by the multilayer perceptron are essentially unbounded real numbers, and the variance representing uncertainty must be strictly positive in mathematical definition, and the direct original numerical values ​​lack clear probabilistic and statistical meaning, a probability distribution model conforming to physical reality cannot be constructed without specific analysis and constraint processing. Therefore, in the technical solution of this application, the original output vector is further analyzed and constrained for probability distribution parameters to obtain the predicted heat load and prediction uncertainty, thereby transforming the abstract output of the neural network into distribution parameters that conform to statistical laws. This ensures that the output prediction results not only have the physical meaning of the predicted expectation mathematically, but also numerically satisfy the domain constraint of the probability density function, thus providing a rigorous and reliable quantitative basis for subsequent risk-aware control decisions.

[0030] Specifically, in one example of this application, the processor of the intelligent temperature control system executes an analytical transformation program from unconstrained raw numerical values ​​to physically meaningful statistical parameters. For the processing of the raw output vector, the system first extracts the first dimension component of the vector and directly maps it to the mean parameter of a Gaussian probability distribution, which represents the predicted heat load. Subsequently, the system extracts the second dimension component of the raw output vector, performs an exponential activation function operation on it to ensure the non-negativity of the value, and then superimposes a preset minimum positive number as a numerical stabilizing term, thereby mapping the calculation result to the variance parameter of a Gaussian probability distribution. Then, the square root of this variance parameter is performed to obtain the standard deviation, which is represented as the prediction uncertainty for dimensionally unified calculation with a thermodynamic buffer space having the dimension of temperature in subsequent steps.

[0031] Specifically, in step S500, based on the prediction uncertainty, prediction-based feedforward-feedback coordinated control is performed on the predicted heat load and the current state to obtain actuator commands. It should be understood that, since relying solely on prediction information for feedforward control may lead to overly aggressive actions due to prediction errors, while relying solely on feedback control suffers from response lag, a key challenge in prediction-based feedforward-feedback coordinated control systems lies in how to integrate uncertain prediction information with reliable real-time feedback. Traditional fusion mechanisms suffer from a technical flaw in that their handling of prediction uncertainty is state-independent. Specifically, the function used to adjust the feedforward control weights in this mechanism takes only the magnitude of the prediction uncertainty as its input variable. This means that regardless of the actual operating state of the system—for example, whether the indoor temperature is at the center of the comfort zone or is approaching the boundary of the comfort zone—as long as the measurement of prediction uncertainty is the same, the weight assigned to the feedforward control remains fixed. This strategy severely deviates from the actual industrial control risk logic. The essence of the problem is that the actual risk brought about by prediction uncertainty has a strongly coupled, nonlinear, special relationship with the current system state (specifically, the distance between the current measured value and the safety boundary). When the system state has sufficient buffer space, even with high uncertainty in the prediction, the system can withstand a certain degree of feedforward control deviation, because even if the control action is slightly excessive, it is unlikely to immediately lead to performance violations. However, when the system state is close to the safety boundary, any small uncertainty may induce serious consequences of control overrun. At this point, even if the value of the prediction uncertainty itself is not large, the potential risk it contains is extremely high. Therefore, existing mechanisms simply statically map the prediction uncertainty value to the feedforward weights, failing to effectively quantify and integrate the control risks inherent in the current state, hindering the system from achieving true risk avoidance and dynamic adaptive control. Its weight adjustment strategy appears too one-size-fits-all, lacking the ability to handle fine-grained and differentiated processing under different system states, especially in critical states near the boundary, where the system will behave recklessly and fragilely. To solve the above problems, this technical solution proposes an adaptive feedforward-feedback fusion control method based on state-dependent risk assessment. That is, in the technical solution of this application, based on prediction uncertainty, prediction-based feedforward-feedback collaborative control is performed on the predicted heat load and the current state to obtain actuator commands. This allows for the construction of a state-dependent risk assessment and dynamic allocation mechanism, which adjusts the aggressiveness of the control strategy in real time based on the current thermodynamic buffer space of the system and the reliability of the prediction. This ensures that the system fully utilizes the energy-saving and pre-adjustment advantages of feedforward control when the safety margin is sufficient and the prediction is reliable, while automatically and smoothly switching to a conservative and robust feedback control mode when approaching the safety boundary or when the prediction is questionable. Thus, it maximizes the performance and efficiency of the control system while ensuring absolute safety.

[0032] Figure 4 This is a flowchart illustrating the process of obtaining actuator commands by performing prediction-based feedforward-feedback coordinated control on the predicted heat load and current state, based on the prediction uncertainty of the intelligent temperature control method based on heat load prediction according to embodiments of this application. Figure 4 As shown, step S500 includes: S510, quantifying the prediction uncertainty, current state, and comfort range using thermodynamic buffer space and dynamic risk factors to obtain risk factors; S520, determining adaptive weights based on risk factors; and S530, synthesizing the feedforward control quantity and feedback control quantity based on risk perception based on the adaptive weights to obtain the actuator command.

[0033] Accordingly, in step S510, the prediction uncertainty, current state, and comfort zone are quantified using thermodynamic buffer space and dynamic risk factors to obtain risk factors. First, thermodynamic buffer space quantification is performed. Specifically, the processor of the intelligent temperature control system executes a state-dependent two-stage risk assessment procedure. First, the system performs quantification calculation of the thermodynamic buffer space based on real-time acquired current state data and preset comfort zone boundaries. Specifically, the system reads the current indoor temperature collected by the sensor in real time and retrieves the upper and lower limits of the preset comfort temperature range. Then, the Euclidean distance from the current temperature point to the nearest comfort boundary is calculated, and this distance is defined as the thermodynamic buffer space at the current moment. The calculation formula for this process is as follows: in, The thermodynamic buffer space at time t. Represents the indoor temperature at time t. and These represent the upper and lower limits of the comfort temperature range, respectively. It should be understood that a simple temperature value cannot directly reflect the safety state of a system; for example, the same temperature fluctuation can pose drastically different risks at the center and edge of the comfort zone. Therefore, in this application's technical solution, the Euclidean distance from the current temperature point to the nearest comfort boundary is further calculated as a measure of the thermodynamic buffer space. This buffer space value represents the system's operational space or safety margin without compromising user comfort; it is a direct quantification of the current system's safety state. The larger the value, the safer the system, and naturally, the higher its tolerance for uncertainty. This provides a crucial state-dependent benchmark for subsequent risk assessment, ensuring that the risk assessment is not only based on the prediction itself but also on the physical environmental context in which the system exists.

[0034] Next, based on the calculated thermodynamic buffer space and the prediction uncertainty of the model output, a comprehensive dynamic risk factor is constructed. Specifically, the system extracts the prediction uncertainty of the prediction model output from the previous step (i.e., the quantified result converted to standard deviation) as the numerator; simultaneously, it extracts the just-calculated thermodynamic buffer space as the denominator, and introduces adjustment coefficients and stability constants to construct a ratio-form function model to calculate the dynamic risk factor. The calculation formula for this process is shown below: in, The dynamic risk factor at time t is represented. It is the standard deviation of the predicted distribution at time t. It is an adjustable risk sensitivity coefficient. This is a tiny normal number used to prevent the denominator from being zero. It should be understood that neither predictive uncertainty nor buffer space alone is sufficient to fully assess risk; both must be combined to capture their coupling effect. Therefore, in the technical solution of this application, the standard deviation of predictive uncertainty is further used as the numerator, and the thermodynamic buffer space as the denominator to construct a ratio-based risk factor, thereby reflecting the logical relationship that risk is directly proportional to uncertainty and inversely proportional to buffer space. Even with a very small predictive uncertainty, if the current buffer space tends to zero, the risk factor will be drastically amplified; conversely, even with high uncertainty, if the buffer space is large enough, the risk factor may still be maintained at a low level. In this way, a dynamic indicator that accurately reflects the current comprehensive risk level can be obtained, thereby guiding subsequent control strategies to adaptively switch between aggressive and conservative approaches.

[0035] Accordingly, in step S520, adaptive weights are determined based on the risk factor. It should be understood that since the dynamic risk factor calculated in the preceding steps is an unbounded physical quantity, its value only represents the relative degree of risk and cannot be directly used as a weighting coefficient for the weighted summation calculation of control quantities. Furthermore, the relationship between the increase in risk and the conservatism of the control strategy is not a simple linear one; a linear mapping cannot achieve rapid response and suppression of high-risk states. Therefore, in order to map the unbounded risk factor into a bounded weight value that can be used for weighting control quantities, the technical solution of this application further determines adaptive weights based on the risk factor, thereby constructing a nonlinear conversion mechanism that transforms abstract risk assessment indicators into specific control authority allocation signals. This ensures that the aggressiveness of the control strategy is sensitively and intelligently adjusted according to changes in risk; that is, under low risk, feedforward control is fully authorized to improve performance, while feedforward authority is rapidly reduced to ensure safety when risk increases, thus realizing a logical closed loop from risk perception to decision execution.

[0036] Specifically, in one example of this application, the processor of the intelligent temperature control system executes a weight generation program based on nonlinear mapping. The system first extracts the dynamic risk factor at time t calculated in the previous stage; this factor numerically belongs to a positive semi-definite infinite interval. Then, the system calls a preset nonlinear exponential decay function model, using the dynamic risk factor as the input variable and combining it with a pre-calibrated decay rate parameter for calculation. This function can respond quickly and smoothly to increases in risk. Through this calculation, the system maps a risk factor with potentially extremely large values ​​to a dimensionless value strictly constrained within a closed interval of zero to one, i.e., generating an adaptive weight between 0 and 1. This adaptive weight directly characterizes the system's level of confidence in the feedforward control strategy and serves as the core adjustment coefficient in the subsequent instruction synthesis step. The calculation formula for this process is shown below: in, Represents the adaptive weights at time t. It is the dynamic risk factor at time t. It is a decay rate parameter representing the rate at which the control weight decreases as risk increases. This formula defines, in practical temperature control scenarios, how the aggressiveness of the control strategy intelligently adjusts with changes in risk: when dynamic risk factors... When the system approaches zero (i.e., the system is in the center of a very safe comfort zone and the prediction is highly reliable), adaptive weights are used. Approaching 1, at which point the system fully trusts and employs efficient feedforward control; with dynamic risk factors... The increase in uncertainty (i.e., increased prediction uncertainty or the system approaching the safety boundary) necessitates adaptive weighting. The parameters decrease exponentially and rapidly, approaching zero, indicating a rapid decline in the system's confidence in feedforward control, and the control strategy automatically degenerates into conservative feedback control. This determines the sensitivity of the degradation. The larger the value, the more sensitive the system is to risk. Even the slightest sign of risk will immediately shut down feedforward control, thus ensuring the system's absolute robustness in critical states. This aims to generate a weighted signal that reflects the risk level in real time and is used for dynamically allocating control permissions, thereby preparing for the final instruction synthesis.

[0037] Accordingly, in step S530, based on adaptive weights, risk-aware instruction synthesis is performed on the feedforward control quantity and the feedback control quantity to obtain the actuator instruction. It should be understood that static superposition of feedforward and feedback cannot adapt to the dynamic changes between prediction reliability and the safety of the current system state, and simple addition cannot guarantee the conservation of control authority, which may lead to actuator saturation or instruction conflicts. Therefore, in the technical solution of this application, risk-aware instruction synthesis is further performed on the feedforward control quantity and the feedback control quantity based on adaptive weights to obtain the actuator instruction, thereby implementing a risk-aware dynamic authority allocation mechanism. Based on the generated adaptive weights, the feedforward and feedback control components are ultimately and intelligently fused. This ensures that the control authority smoothly and logically shifts between aggressive prediction strategies and conservative feedback strategies according to the real-time assessed risk level, thus guaranteeing that the final output instruction represents the optimal trade-off between performance and safety.

[0038] More specifically, in a specific example of this application, the processor of the intelligent temperature control system executes a dual-channel control component calculation and final instruction synthesis program. First, the feedforward control quantity is calculated in the feedforward channel: by multiplying the predicted heat load output by the prediction model by a preset conversion gain coefficient, the heat load demand value is converted into the corresponding valve opening or power increment, thus obtaining the feedforward control quantity. Simultaneously, the feedback control quantity is calculated in the feedback channel: by inputting the current indoor temperature and the preset temperature setpoint into a standard PID controller, the deviation compensation value is calculated through proportional, integral, and derivative operations, thus obtaining the feedback control quantity. Subsequently, the adaptive weights generated in the preceding steps are... The feedforward and feedback control variables are subjected to a risk-aware-based convex combination operation to generate the final actuator instruction. This synthesis method differs fundamentally from simple addition; it embodies the dynamic transfer of control authority between the two control strategies while ensuring the conservation of total control authority. The calculation formula for this process is shown below: in, This represents the executor instruction finally generated at time t. It is the adaptive weight at time t. It is a feedforward control variable. It is a feedback control variable. Its contextual significance lies in the fact that when the risk is extremely low ( Approaching 1), the final instructions will mainly consist of the feedforward part, giving the system a strong sense of foresight and initiative; however, when the risk is extremely high ( (Approaching 0), the final instruction will consist almost entirely of the feedback part. At this point, the system will automatically degenerate into a robust and stable traditional PID controller, thus ensuring absolute safety in the most uncertain critical state.

[0039] Through the above preferred embodiments, an intelligent control system capable of seamlessly and smoothly switching between aggressive predictive and conservative corrective control modes is achieved. The switching decision is no longer based solely on predictive uncertainty, but rather on a comprehensive risk assessment combining the uncertainty of future predictive information with the safety of the current system state. In this way, the system can maximize the use of high-precision multimodal predictive information to make advance adjustments while ensuring operational safety and meeting the hard constraints of user comfort, thereby significantly improving building energy efficiency and occupant comfort. This method ultimately ensures that the control system exhibits excellent robustness and intelligence under all operating conditions, especially in challenging scenarios close to performance boundaries.

[0040] In summary, the intelligent temperature control method based on heat load prediction according to the embodiments of this application is explained. It first performs modal-specific extraction and cross-modal feature fusion on time-series and image data streams to construct a multimodal representation, simultaneously outputting the predicted heat load and uncertainty quantification results. Based on this, by coupling the current thermodynamic buffer space with the prediction uncertainty, a dynamic risk factor is calculated, and then the combined weights of feedforward and feedback control are adaptively adjusted. This mechanism maximizes the energy-saving advantages of feedforward regulation while ensuring a safety baseline, effectively solving the technical problems of inaccurate single-modal prediction and poor robustness of static control strategies.

[0041] Furthermore, an intelligent temperature control system based on heat load prediction is also provided.

[0042] Figure 5 This is a block diagram of an intelligent temperature control system based on heat load prediction according to an embodiment of this application. Figure 5 As shown, the intelligent temperature control system 100 based on heat load prediction according to an embodiment of this application includes: a real-time data acquisition module 110 for acquiring time-series data streams and image data streams; a modality-specific feature extraction module 120 for extracting modality-specific features from the time-series data streams and image data streams to obtain time-series sequence embeddings and visual embeddings; a multimodal feature fusion module 130 for fusing features from the time-series sequence embeddings and visual embeddings to obtain a multimodal fusion representation vector; a heat load and uncertainty prediction module 140 for predicting heat load and quantifying uncertainty in the multimodal fusion representation vector to obtain predicted heat load and predicted uncertainty; and a feedforward-feedback collaborative control module 150 for performing prediction-based feedforward-feedback collaborative control on the predicted heat load and current state based on the predicted uncertainty to obtain actuator commands.

[0043] As described above, the intelligent temperature control system 100 based on heat load prediction according to the embodiments of this application can be implemented in various wireless terminals, such as servers with intelligent temperature control algorithms based on heat load prediction. In one possible implementation, the intelligent temperature control system 100 based on heat load prediction according to the embodiments of this application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the intelligent temperature control system 100 based on heat load prediction can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the intelligent temperature control system 100 based on heat load prediction can also be one of many hardware modules of the wireless terminal.

[0044] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. An intelligent temperature control method based on thermal load prediction, characterized in that, The method comprises the following steps: obtaining a time series data stream and an image data stream; performing modal-specific feature extraction on the time series data stream and the image data stream to obtain a time series sequence embedding and a visual embedding; performing feature fusion on the time series sequence embedding and the visual embedding to obtain a multi-modal fusion representation vector; performing thermal load prediction and uncertainty quantification on the multi-modal fusion representation vector to obtain a predicted thermal load and a predicted uncertainty; based on the predicted uncertainty, performing a prediction-based feedforward-feedback collaborative control on the predicted thermal load and a current state to obtain an actuator instruction. 2.The intelligent temperature control method based on thermal load prediction of claim 1, wherein, The modal-specific feature extraction on the time series data stream and the image data stream to obtain the time series sequence embedding and the visual embedding comprises: inputting the time series data stream into a time series encoder to obtain the time series sequence embedding; inputting the image data stream into an image feature encoder to obtain the visual embedding. 3.The intelligent temperature control method based on thermal load prediction of claim 2, wherein, The time series encoder is a one-dimensional convolutional neural network model, and the image feature encoder is a pre-trained two-dimensional convolutional neural network model. 4.The intelligent temperature control method based on thermal load prediction of claim 1, wherein, The feature fusion on the time series sequence embedding and the visual embedding to obtain the multi-modal fusion representation vector comprises: performing intra-modal encoding on the time series sequence embedding and the visual embedding to obtain a time series intra-modal correlation encoding vector and a visual intra-modal correlation feature vector; performing cross-modal encoding on the time series intra-modal correlation encoding vector and the visual intra-modal correlation feature vector to obtain the multi-modal fusion representation vector. 5.The intelligent temperature control method based on thermal load prediction of claim 1, wherein, The thermal load prediction and uncertainty quantification on the multi-modal fusion representation vector to obtain the predicted thermal load and the predicted uncertainty comprises: inputting the multi-modal fusion representation vector into a multi-layer perceptron for regression prediction to obtain an original output vector; performing probability distribution parameter analysis and constraint on the original output vector to obtain the predicted thermal load and the predicted uncertainty. 6.The intelligent temperature control method based on thermal load prediction of claim 1, wherein, The prediction-based feedforward-feedback collaborative control on the predicted thermal load and the current state based on the predicted uncertainty to obtain the actuator instruction comprises: performing thermodynamic buffer space and dynamic risk factor quantification on the predicted uncertainty, the current state and a comfort interval to obtain a risk factor; based on the risk factor, determining an adaptive weight; based on the adaptive weight, performing risk-aware instruction synthesis on a feedforward control quantity and a feedback control quantity to obtain the actuator instruction. 7.The intelligent temperature control method based on thermal load prediction of claim 6, wherein, The determination process of the feedforward control quantity and the feedback control quantity comprises: multiplying the predicted thermal load by a conversion gain coefficient to obtain the feedforward control quantity; inputting an indoor temperature in the current state and a temperature set point into a PID controller to obtain the feedback control quantity. 8.The intelligent temperature control method based on thermal load prediction of claim 6, wherein, The thermodynamic buffer space and dynamic risk factor quantification on the predicted uncertainty, the current state and the comfort interval to obtain the risk factor comprises: performing thermodynamic buffer space and dynamic risk factor quantification on the predicted uncertainty, the current state and the comfort interval according to the following formula: wherein, represents the thermodynamic buffer space at time t, is a minimum function, represents the indoor temperature at time t, and represent the upper and lower limits of the comfort temperature interval, respectively, represents the dynamic risk factor at time t, is the standard deviation of the predicted distribution at time t, is an adjustable risk sensitivity coefficient, is a very small positive number used to prevent division by zero.

9. An intelligent temperature control system based on thermal load prediction, characterized in that, The method comprises the following steps: a real-time data acquisition module for obtaining a time series data stream and an image data stream; a modal-specific feature extraction module for performing modal-specific feature extraction on the time series data stream and the image data stream to obtain a time series sequence embedding and a visual embedding; a multi-modal feature fusion module for performing feature fusion on the time series sequence embedding and the visual embedding to obtain a multi-modal fusion representation vector; a thermal load and uncertainty prediction module configured to perform thermal load prediction and uncertainty quantification on the multi-modal fusion representation vector to obtain a predicted thermal load and a predicted uncertainty; a feedforward-feedback collaborative control module configured to perform a prediction-based feedforward-feedback collaborative control on the predicted thermal load and the current state based on the predicted uncertainty to obtain an actuator instruction.

Images