Perceptual CNN-LSTM hybrid prediction control method and system based on Bayesian optimization
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-23
- Publication Date
- 2026-03-31
AI Technical Summary
In the optimization control of air source heat pump (ASHP)-phase change material (PCM) systems, existing technologies rely on manual trial and error for hyperparameters, lacking a systematic optimization mechanism. Furthermore, the model training does not consider the economic cost differences in prediction errors at different times under time-of-use pricing, resulting in insufficient prediction accuracy and economy.
The hyperparameters of the CNN-LSTM hybrid prediction model are adaptively optimized using the Bayesian optimization algorithm and trained in combination with the electricity price weighted loss function to achieve cost-aware control and improve prediction accuracy and economy.
Through adaptive optimization and cost-aware training, the model's prediction accuracy and operational economy have been improved, enabling the ASHP-PCM system to operate in a "peak-shifting and valley-filling" manner, alleviating the pressure of peak-valley load difference in the power grid and reducing user costs.
Smart Images
Figure CN121763775A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of building energy optimization and control technology, and in particular to a perceptual CNN-LSTM hybrid predictive control method and system based on Bayesian optimization. Background Technology
[0002] Against the backdrop of global energy transition and the advancement of carbon neutrality goals, building electrification has become a key path to reduce dependence on fossil fuels and decrease carbon emissions. Air source heat pumps (ASHPs), with their high energy efficiency and low pollution characteristics, are gradually replacing traditional coal-fired and gas-fired heating equipment, becoming the mainstream choice for building heating. However, the centralized connection of large-scale ASHPs to the power grid significantly exacerbates the peak-valley load difference in the power grid—especially during peak heating seasons, the simultaneous operation of numerous ASHPs leads to a surge in grid load. Furthermore, the increasing penetration of renewable energy further exacerbates the imbalance between power grid supply and demand, seriously affecting the safe and stable operation of the power grid.
[0003] To alleviate the aforementioned problems, integrated phase change material (PCM) thermal energy storage systems are widely used in air source heat pump (ASHP) systems. PCMs can utilize latent heat to achieve efficient energy storage and release, offering advantages such as high heat storage density and low temperature fluctuations. This enables peak shaving and valley filling of electricity demand, relieving grid pressure and reducing user operating costs. However, the optimized control of ASHP-PCM coupled systems faces numerous challenges: the heat absorption and release processes of PCMs involve complex solid-liquid phase change interface movements, exhibiting strong nonlinearity and large thermal inertia. While traditional physics-based modeling methods offer high accuracy, they are computationally intensive, rely on precise thermophysical parameters, and struggle to meet real-time predictive control requirements.
[0004] Data-driven models, due to their lack of reliance on complex physical mechanisms and high computational efficiency, have become a research hotspot in predictive control of ASHP-PCM systems. Convolutional Neural Networks (CNNs) excel at extracting multivariate spatial coupling features, while Long Short-Term Memory Networks (LSTMs) effectively capture long-range dependencies in time-series data. Hybrid models combining these two technologies exhibit excellent performance in load forecasting. However, existing technologies still have significant drawbacks: firstly, the hyperparameters of hybrid models often rely on manual trial and error or grid search, lacking a systematic optimization mechanism and making it difficult to achieve optimal model performance; secondly, model training often focuses on minimizing statistical errors, neglecting the economic cost differences in prediction errors across different time periods under time-of-use pricing, resulting in models that are statistically "accurate" but economically "non-optimal." Therefore, to better address these issues, this invention proposes a perceptual CNN-LSTM hybrid predictive control method and system based on Bayesian optimization. Summary of the Invention
[0005] To address the above issues, this invention proposes a perceptual CNN-LSTM hybrid predictive control method and system based on Bayesian optimization. By using Bayesian optimization to achieve adaptive hyperparameter optimization, and combining it with an electricity price-weighted loss function to construct a cost-aware training mechanism, the prediction accuracy and operational economy are improved.
[0006] A perceptual CNN-LSTM hybrid predictive control method based on Bayesian optimization includes the following steps:
[0007] Step S1: Digital modeling of physical objects and generation of multi-source datasets: A high-fidelity simulation model of the air source heat pump (ASHP)-phase change material (PCM) coupled system is established based on TRNSYS, and a multi-dimensional time series dataset containing meteorological parameters, building thermal parameters and system operating status parameters is collected and generated.
[0008] Step S2: Multi-source time series data preprocessing: Normalize, slide window truncation and time series periodic encoding are performed on the multi-dimensional time series dataset generated in step S1 to obtain a standardized input sequence, and predict the system heat load in the next 24 hours.
[0009] Step S3: Construct a cost-aware CNN-LSTM hybrid prediction model with integrated Bayesian optimization: The hybrid prediction model includes a CNN feature extraction branch, an LSTM temporal modeling branch, and an output layer; CNN is used to extract the spatial coupling features between multiple variables in the standardized input sequence, LSTM is used to capture the long-term thermal inertia features of the PCM phase transition process, and the output layer fuses the features to obtain the thermal load prediction result;
[0010] Step S4: Adaptive optimization of model hyperparameters based on Bayesian optimization: Using the Bayesian optimization algorithm and Gaussian process regression as a surrogate model, the optimal combination of hyperparameters of the CNN-LSTM hybrid prediction model is automatically searched and determined. The hyperparameters include the kernel size, the number of hidden units in the LSTM layer, and the learning rate.
[0011] Step S5: Cost-aware model training: Introduce an electricity price-weighted loss function. During the model training process, dynamically adjust the weight coefficients of the prediction error at different times based on the time-of-use electricity price signal. The weight coefficients for peak electricity price periods are greater than those for off-peak electricity price periods. Minimize the weighted loss value through the gradient descent algorithm to complete the model training.
[0012] Step S6: Predictive guidance for ASHP-PCM system optimization control: Utilize the trained prediction model to output the predicted heat load for the next 24 hours, and formulate a control strategy based on the time-of-use electricity price signal. Control ASHP to operate during off-peak hours and store heat in the PCM energy storage device, and control ASHP to stop operating during peak hours and release heat from the PCM energy storage device for heating.
[0013] Furthermore, in step S1, the parameters of the multidimensional time-series dataset include outdoor temperature, indoor temperature, solar radiation intensity, ASHP operating power, PCM temperature, PCM thermal storage status, historical electricity price, and historical heat load.
[0014] Furthermore, in step S2, the multi-source time series data preprocessing includes normalization of the multi-dimensional time series dataset, sliding window truncation, and time series periodic encoding. The sliding window has a time length of 24 hours and is used to predict the system heat load for the next 24 hours using the past 24 hours of data as input. The time series periodic encoding uses sine / cosine transform to encode the timestamps to characterize the continuity of the time dimension.
[0015] Furthermore, in step S3, the CNN feature extraction branch includes a 1D convolutional layer, a batch normalization layer, and a ReLU activation function layer. The 1D convolutional layer slides along the time axis to extract local temporal features. The LSTM temporal modeling branch also includes a Dropout layer to prevent model overfitting.
[0016] Further, in step S5, the expression for the electricity price-weighted loss function is: In the formula, L is the weighted loss value, and N is the total number of prediction time steps. Let be the electricity price weighting coefficient at time t. Let be the actual heat load value at time t. Let be the predicted heat load value at time t. It is positively correlated with the time-of-use electricity price at time t.
[0017] A Bayesian-optimized perceptual CNN-LSTM hybrid predictive control system includes a physical system layer, a data processing layer, and an intelligent prediction and decision-making layer.
[0018] Physical system layer: includes air source heat pump unit, phase change energy storage tank, building terminal load unit and environmental and power grid sensing unit, the environmental and power grid sensing unit is used to collect meteorological parameters and time-of-use electricity price signals;
[0019] Data processing layer: Used to receive multi-source data collected by the physical system layer, perform normalization, sliding window truncation and time-series periodic encoding processing, and output a standardized input sequence;
[0020] Intelligent prediction and decision-making layer: including CNN-LSTM hybrid prediction module, Bayesian optimization module, cost-aware training module and optimization control module;
[0021] The Bayesian optimization module is used to provide the optimal hyperparameter combination for the CNN-LSTM hybrid prediction module; the cost-aware training module trains the CNN-LSTM hybrid prediction module using an electricity price-weighted loss function; the CNN-LSTM hybrid prediction module is used to output the predicted heat load value for the next 24 hours; the optimization control module generates control commands based on the predicted heat load value and the time-of-use electricity price signal to control the operating status of the air source heat pump unit and the phase change energy storage tank in the physical system layer.
[0022] Furthermore, the environment and power grid sensing unit includes a temperature sensor, a radiation sensor, and an electricity price signal receiving module, which are used to collect outdoor temperature, solar radiation intensity, and time-of-use electricity price signals, respectively.
[0023] Furthermore, the intelligent prediction and decision-making layer also includes a data storage module for storing the raw data collected by the physical system layer, the standardized data output by the data processing layer, and the intermediate data during the model training process.
[0024] Compared with the prior art, the significant advantages of this invention are:
[0025] (1) This invention uses Bayesian optimization algorithm to achieve adaptive optimization of hyperparameters of CNN-LSTM hybrid prediction model. Gaussian process regression is used as a surrogate model. The optimal hyperparameter combination is searched efficiently through the Acquisition function. This method abandons the traditional manual trial and error and grid search method, reduces the blindness of hyperparameter optimization, improves the model optimization efficiency and prediction performance, and ensures that the model achieves the best prediction accuracy.
[0026] (2) This invention innovatively introduces a price-weighted loss function, which directly correlates the time-of-use price signal with the prediction error and dynamically adjusts the weight coefficient of the prediction error at different times. This allows the model to prioritize the optimization of the prediction accuracy during peak electricity price periods during the training process, reducing the high economic cost caused by the prediction error during peak periods. This realizes the transformation of the model from "statistically accurate" to "economically optimal", and significantly improves the operational economy of the ASHP-PCM system.
[0027] (3) The CNN-LSTM hybrid prediction model constructed in this invention extracts the spatial coupling features between multiple variables through CNN and captures the long-term thermal inertia features of the PCM phase transition process through LSTM, thus realizing the effective fusion of spatiotemporal features. It can accurately capture the nonlinear dynamic characteristics of the ASHP-PCM coupled system and provide a reliable prediction basis for system optimization control.
[0028] (4) The predictive control method and system of the present invention can realize the "peak shifting and valley filling" operation of the ASHP-PCM system, which stores heat during the valley period and releases heat during the peak period. This not only alleviates the pressure of the peak-valley load difference of the power grid, but also reduces the operating costs of users. It has significant economic and social benefits and broad application prospects. Attached Figure Description
[0029] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:
[0030] Figure 1 This is a flowchart of a perceptual CNN-LSTM hybrid predictive control method based on Bayesian optimization according to the present invention. Detailed Implementation
[0031] middle
[0032] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.
[0033] A perceptual CNN-LSTM hybrid predictive control method based on Bayesian optimization includes the following steps:
[0034] Step S1: Digital modeling of physical objects and generation of multi-source datasets: A high-fidelity simulation model of the air source heat pump (ASHP)-phase change material (PCM) coupled system is established based on TRNSYS, and a multi-dimensional time series dataset containing meteorological parameters, building thermal parameters and system operating status parameters is collected and generated.
[0035] In specific embodiments, such as Figure 1 As shown, a high-fidelity simulation model of the ASHP-PCM coupled system was established based on TRNSYS software. In the model, the ASHP unit is an air source heat pump with a rated power of 10kW, and the nonlinear characteristics of COP changing with outdoor temperature were obtained by fitting experimental data. The PCM uses paraffin-based phase change material with a phase change temperature of 28℃, filled in a 500L energy storage tank. The building terminal load is designed based on the heating load of a 100㎡ residential building. Multidimensional time-series data for one year were collected through the simulation model, with a sampling interval of 1 hour, generating a multidimensional time-series dataset including outdoor temperature, indoor temperature, solar radiation intensity, ASHP operating power, PCM temperature, PCM state of heat storage (SoC), historical electricity price, and historical heat load.
[0036] Step S2: Multi-source time series data preprocessing: Normalize, slide window truncation and time series periodic encoding are performed on the multi-dimensional time series dataset generated in step S1 to obtain a standardized input sequence, and predict the system heat load in the next 24 hours.
[0037] In specific embodiments, such as Figure 1 As shown, the multidimensional time-series dataset generated in step S1 is normalized, subjected to sliding window truncation, and time-series periodic encoding to obtain a standardized input sequence. The time length of the sliding window is 24 hours, which is used to predict the system heat load for the next 24 hours using the past 24 hours of data as input. The time-series periodic encoding uses sine / cosine transform to encode the timestamp to characterize the continuity of the time dimension. For example, time t is encoded by sin(2πt / T) and cos(2πt / T), where T is the time period.
[0038] Step S3: Construct a cost-aware CNN-LSTM hybrid prediction model with integrated Bayesian optimization: The hybrid prediction model includes a CNN feature extraction branch, an LSTM temporal modeling branch, and an output layer; CNN is used to extract the spatial coupling features between multiple variables in the standardized input sequence, LSTM is used to capture the long-term thermal inertia features of the PCM phase transition process, and the output layer fuses the features to obtain the thermal load prediction result;
[0039] In specific embodiments, such as Figure 1 As shown, the hybrid prediction model includes a CNN feature extraction branch, an LSTM temporal modeling branch, and an output layer. The CNN feature extraction branch includes a 1D convolutional layer, a batch normalization layer, and a ReLU activation function layer. The 1D convolutional layer slides along the time axis to extract local temporal features, the batch normalization layer is used to accelerate model convergence, and the ReLU activation function layer is used to introduce nonlinear features. The LSTM temporal modeling branch includes an LSTM layer and a Dropout layer. The LSTM layer is used to capture the long-term thermal inertia features of the PCM phase transition process, and the Dropout layer is used to prevent the model from overfitting. The output layer uses a fully connected layer and a regression layer. The thermal load prediction result is obtained by fusing the spatial features extracted by the CNN and the temporal features captured by the LSTM.
[0040] Step S4: Adaptive optimization of model hyperparameters based on Bayesian optimization: Using the Bayesian optimization algorithm and Gaussian process regression as a surrogate model, the optimal combination of hyperparameters of the CNN-LSTM hybrid prediction model is automatically searched and determined. The hyperparameters include the kernel size, the number of hidden units in the LSTM layer, and the learning rate.
[0041] In specific embodiments, such as Figure 1 As shown, a Bayesian optimization algorithm is used, with Gaussian process regression as a surrogate model, to automatically search for and determine the optimal hyperparameter combination of the CNN-LSTM hybrid prediction model. The hyperparameters include the kernel size, the number of hidden units in the LSTM layer, and the learning rate. The specific optimization process includes:
[0042] S41: Initialize the hyperparameter search space and set the value range of each hyperparameter;
[0043] S42: Randomly sample a set of initial hyperparameters within the search space, train the CNN-LSTM hybrid prediction model, and calculate the validation set error;
[0044] S43: Construct a surrogate model using Gaussian process regression, and infer the performance of unsampled hyperparameter combinations based on the validation set error of sampled hyperparameters;
[0045] S44: Select the next set of hyperparameters to be sampled using the Acquisition function (e.g., to improve the EI function);
[0046] S45: Repeat steps S42-S44 until the preset number of iterations is reached or the validation set error converges, and output the optimal hyperparameter combination.
[0047] Step S5: Cost-aware model training: Introduce an electricity price-weighted loss function. During the model training process, dynamically adjust the weight coefficients of the prediction error at different times based on the time-of-use electricity price signal. The weight coefficients for peak electricity price periods are greater than those for off-peak electricity price periods. Minimize the weighted loss value through the gradient descent algorithm to complete the model training.
[0048] In specific embodiments, such as Figure 1 As shown, a price-weighted loss function is introduced. During model training, the weight coefficients of the prediction error at different times are dynamically adjusted according to the time-of-use electricity price signal. The weight coefficients corresponding to peak electricity price periods are greater than those for off-peak electricity price periods. The weighted loss value is minimized through the gradient descent algorithm to complete model training. The expression of the price-weighted loss function is: , where L is the weighted loss value, N is the total number of prediction time steps, is the electricity price weight coefficient at time t, is the actual heat load value at time t, and is the predicted heat load value at time t, which is positively correlated with the time-of-use electricity price at time t. Specifically, it can be achieved by = k·, where k is the proportionality coefficient, and is the time-of-use electricity price at time t.
[0049] Step S6: Predictive guidance for ASHP-PCM system optimization control: Utilize the trained prediction model to output the predicted heat load for the next 24 hours, and formulate a control strategy based on the time-of-use electricity price signal. Control ASHP to operate during off-peak hours and store heat in the PCM energy storage device, and control ASHP to stop operating during peak hours and release heat from the PCM energy storage device for heating.
[0050] In specific embodiments, such as Figure 1As shown, the predicted heat load for the next 24 hours is output using the trained prediction model. Combined with the time-of-use electricity price signal, a control strategy is formulated. During off-peak hours, the ASHP is controlled to operate and store heat in the PCM energy storage device. During peak hours, the ASHP is controlled to stop operating and the PCM energy storage device releases heat for heating, thereby achieving peak shifting and valley filling while minimizing operating costs.
[0051] A Bayesian-optimized perceptual CNN-LSTM hybrid predictive control system includes a physical system layer, a data processing layer, and an intelligent prediction and decision-making layer.
[0052] Physical system layer: includes air source heat pump unit, phase change energy storage tank, building terminal load unit and environmental and power grid sensing unit, the environmental and power grid sensing unit is used to collect meteorological parameters and time-of-use electricity price signals;
[0053] Data processing layer: Used to receive multi-source data collected by the physical system layer, perform normalization, sliding window truncation and time-series periodic encoding processing, and output a standardized input sequence;
[0054] Intelligent prediction and decision-making layer: including CNN-LSTM hybrid prediction module, Bayesian optimization module, cost-aware training module and optimization control module;
[0055] The Bayesian optimization module is used to provide the optimal hyperparameter combination for the CNN-LSTM hybrid prediction module; the cost-aware training module trains the CNN-LSTM hybrid prediction module using an electricity price-weighted loss function; the CNN-LSTM hybrid prediction module is used to output the predicted heat load value for the next 24 hours; the optimization control module generates control commands based on the predicted heat load value and the time-of-use electricity price signal to control the operating status of the air source heat pump unit and the phase change energy storage tank in the physical system layer.
[0056] In a specific embodiment, the physical system layer includes an air source heat pump unit, a phase change energy storage tank, a building terminal load unit, and an environmental and power grid sensing unit. The air source heat pump unit is used to provide a heat source for heating, the phase change energy storage tank is used to store and release heat energy, the building terminal load unit is used to meet the building heating and domestic hot water needs, and the environmental and power grid sensing unit includes a temperature sensor, a radiation sensor, and an electricity price signal receiving module, which are used to collect outdoor temperature, solar radiation intensity, and time-of-use electricity price signals, respectively.
[0057] The data processing layer receives multi-source data collected by the physical system layer, performs normalization, sliding window truncation, and time-series periodic encoding processing, and outputs a standardized input sequence. The data processing layer uses a microcontroller or industrial control computer to realize real-time data processing.
[0058] The intelligent prediction and decision-making layer includes a CNN-LSTM hybrid prediction module, a Bayesian optimization module, a cost-aware training module, and an optimization control module, as well as a data storage module. The data storage module stores the raw data collected by the physical system layer, the standardized data output by the data processing layer, and the intermediate data during model training. The Bayesian optimization module provides the optimal hyperparameter combination for the CNN-LSTM hybrid prediction module. The cost-aware training module trains the CNN-LSTM hybrid prediction module using an electricity price-weighted loss function. The CNN-LSTM hybrid prediction module outputs the predicted heat load for the next 24 hours. The optimization control module generates control commands based on the predicted heat load and the time-of-use electricity price signal to control the operating status of the air source heat pump unit and the phase change energy storage tank in the physical system layer.
[0059] The working process of this invention is as follows:
[0060] The multi-source data collected by the physical system layer is processed by the data processing layer and then input into the CNN-LSTM hybrid prediction module of the intelligent prediction and decision layer. Under the optimal hyperparameters determined by the Bayesian optimization module and the training strategy of the cost-aware training module, this module outputs the predicted heat load value for the next 24 hours. The optimization control module controls the operation of the physical system layer equipment through the PLC controller based on the predicted value and the time-of-use electricity price signal to achieve the optimization control objective.
[0061] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0062] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.
[0063] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.
Claims
1. A perceptual CNN-LSTM hybrid predictive control method based on Bayesian optimization, characterized in that, Includes the following steps: Step S1: Digital modeling of physical objects and generation of multi-source datasets: A high-fidelity simulation model of the air source heat pump-phase change material coupled system is established based on TRNSYS, and a multi-dimensional time series dataset is collected and generated; Step S2: Multi-source time series data preprocessing: Process the multi-dimensional time series dataset generated in step S1 to obtain a standardized input sequence and predict the system heat load for the next 24 hours; Step S3: Construct a cost-aware CNN-LSTM hybrid prediction model with integrated Bayesian optimization: The hybrid prediction model includes a CNN feature extraction branch, an LSTM temporal modeling branch, and an output layer; CNN is used to extract the spatial coupling features between multiple variables in the standardized input sequence, LSTM is used to capture the long-term thermal inertia features of the PCM phase transition process, and the output layer fuses the features to obtain the thermal load prediction result; Step S4: Adaptive optimization of model hyperparameters based on Bayesian optimization: Using the Bayesian optimization algorithm and Gaussian process regression as a surrogate model, the optimal combination of hyperparameters of the CNN-LSTM hybrid prediction model is automatically searched and determined. Step S5: Cost-aware model training: Introduce an electricity price-weighted loss function. During the model training process, dynamically adjust the weight coefficients of the prediction error at different times based on the time-of-use electricity price signal. The weight coefficients for peak electricity price periods are greater than those for off-peak electricity price periods. Minimize the weighted loss value through the gradient descent algorithm to complete the model training. Step S6: Predictive guidance for ASHP-PCM system optimization control: Utilize the trained prediction model to output the predicted heat load for the next 24 hours, and formulate a control strategy based on the time-of-use electricity price signal. Control ASHP to operate during off-peak hours and store heat in the PCM energy storage device, and control ASHP to stop operating during peak hours and release heat from the PCM energy storage device for heating.
2. The perceptual CNN-LSTM hybrid predictive control method based on Bayesian optimization according to claim 1, characterized in that, in step S1, the parameters of the multidimensional time series dataset include outdoor temperature, indoor temperature, solar radiation intensity, ASHP operating power, PCM temperature, PCM thermal storage status, historical electricity price, and historical heat load.
3. The perceptual CNN-LSTM hybrid predictive control method based on Bayesian optimization according to claim 1, characterized in that, In step S2, the multi-source time series data preprocessing includes normalization of the multi-dimensional time series dataset, sliding window truncation, and time series periodic encoding. The sliding window has a time length of 24 hours and is used to predict the system heat load for the next 24 hours using the past 24 hours of data as input. The time series periodic encoding uses sine / cosine transform to encode the timestamps to characterize the continuity of the time dimension.
4. The perceptual CNN-LSTM hybrid predictive control method based on Bayesian optimization according to claim 1, characterized in that, In step S3, the CNN feature extraction branch includes a 1D convolutional layer, a batch normalization layer, and a ReLU activation function layer. The 1D convolutional layer slides along the time axis to extract local temporal features. The LSTM temporal modeling branch also includes a Dropout layer to prevent model overfitting.
5. The perceptual CNN-LSTM hybrid predictive control method based on Bayesian optimization according to claim 1, characterized in that, In step S5, the expression for the electricity price-weighted loss function is: In the formula, L is the weighted loss value, and N is the total number of prediction time steps. Let be the electricity price weighting coefficient at time t. Let be the actual heat load value at time t. Let be the predicted heat load value at time t. It is positively correlated with the time-of-use electricity price at time t.
6. A Bayesian-optimized perceptual CNN-LSTM hybrid predictive control system, characterized in that, It includes the physical system layer, the data processing layer, and the intelligent prediction and decision-making layer; Physical system layer: includes air source heat pump unit, phase change energy storage tank, building terminal load unit and environmental and power grid sensing unit, the environmental and power grid sensing unit is used to collect meteorological parameters and time-of-use electricity price signals; Data processing layer: Used to receive multi-source data collected by the physical system layer, perform normalization, sliding window truncation and time-series periodic encoding processing, and output a standardized input sequence; Intelligent prediction and decision-making layer: including CNN-LSTM hybrid prediction module, Bayesian optimization module, cost-aware training module and optimization control module; The Bayesian optimization module is used to provide the optimal hyperparameter combination for the CNN-LSTM hybrid prediction module; the cost-aware training module trains the CNN-LSTM hybrid prediction module using an electricity price-weighted loss function; the CNN-LSTM hybrid prediction module is used to output the predicted heat load value for the next 24 hours; the optimization control module generates control commands based on the predicted heat load value and the time-of-use electricity price signal to control the operating status of the air source heat pump unit and the phase change energy storage tank in the physical system layer.
7. The integrated Bayesian optimization perceptual CNN-LSTM hybrid predictive control system according to claim 6, characterized in that, The environment and power grid sensing unit includes a temperature sensor, a radiation sensor, and an electricity price signal receiving module, which are used to collect outdoor temperature, solar radiation intensity, and time-of-use electricity price signals, respectively.
8. A Bayesian-optimized perceptual CNN-LSTM hybrid predictive control system according to claim 6, characterized in that, The intelligent prediction and decision-making layer also includes a data storage module for storing raw data collected by the physical system layer, standardized data output by the data processing layer, and intermediate data during model training.