Residential integrated energy system energy consumption prediction and intelligent management and control method based on artificial intelligence

Abstract

The invention discloses a residential integrated energy system energy consumption prediction and intelligent management and control method based on artificial intelligence, and the method comprises the steps: achieving the multi-feature fusion load prediction through the comprehensive consideration of the influences of meteorological features and time sequence features, such as environment temperature, relative humidity, solar radiation, and the like, through a deep neural network; establishing a load feasible region based on the multi-order RC model in combination with load prediction and building thermal inertia and thermal comfort requirements; reinforcement learning is combined with a mathematical programming / heuristic algorithm, and residence energy thermal inertia and thermal comfort personalized requirements and multi-source coupling optimization regulation and control are comprehensively regulated and controlled. According to the method, personalized thermal comfort and energy consumption requirements of the user can be considered, real-time performance and robustness are optimized under high-precision load supply, and personalized requirements of different house types and equipment combinations can be met.

Description

Technical Field

[0001] This invention belongs to the field of building energy conservation technology, and relates to thermal energy prediction and optimization control of residential integrated energy systems, and particularly to an artificial intelligence-based method for predicting and intelligently managing energy consumption in residential integrated energy systems. Background Technology

[0002] The building sector accounts for a significant proportion of energy demand, making energy conservation and carbon reduction an urgent task. Integrated energy systems, with electricity as the hub, couple cooling, heating, electricity, and renewable energy sources to achieve multi-energy complementarity and coordinated energy supply, have become an effective way to improve building energy efficiency and reliability. However, multi-source coupling and multi-temporal-scale linkage also significantly increase the complexity of system operation, placing higher demands on building-side heat load prediction and energy optimization.

[0003] On the one hand, building thermal inertia and thermal comfort can be viewed as "virtual energy storage" for peak shaving and flexible regulation. However, due to the fluctuations in renewable energy and the uncertainties of weather and behavior, day-ahead and intraday heat load forecasts must take into account both strong nonlinearity and abrupt changes. Existing statistical and simple machine learning methods are insufficient in characterizing complex spatiotemporal features, and the forecast stability and accuracy are difficult to meet the requirements of engineering applications. On the other hand, current energy optimization methods mostly rely on mathematical programming or heuristic algorithms with preset scenarios and static parameters. Faced with forecast errors and external environmental disturbances, they need to be frequently resolved, resulting in high computational costs and insufficient real-time performance and robustness. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the present invention aims to provide an artificial intelligence-based method for predicting and intelligently managing energy consumption in residential integrated energy systems. This method employs multiple deep learning networks to comprehensively consider the influence of environmental factors and time series data, achieving multi-feature fusion load prediction and improving prediction accuracy. Furthermore, it utilizes reinforcement learning combined with mathematical programming / heuristic algorithms to comprehensively regulate residential energy consumption thermal inertia and thermal comfort, as well as multi-source coupling optimization control, to meet users' personalized thermal comfort, energy saving, and cost-saving needs.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An artificial intelligence-based method for predicting and intelligently managing energy consumption in a residential integrated energy system includes the following steps: By performing multidimensional correlation modeling and pattern extraction of meteorological and time series features, and based on correlation screening and multi-scale feature fusion, the predicted results of different user load demands are obtained. A multi-order RC building thermodynamic model is adopted, and the building thermal inertia, user thermal comfort requirements and the prediction results are combined to generate a dynamic feasible load domain that characterizes the individual indoor state constraints. Considering users' personalized needs for thermal comfort, energy saving, and cost reduction, a two-layer collaborative optimization algorithm is constructed to develop energy-saving and cost-saving strategies that take into account building thermal inertia, user thermal comfort, and multi-source equipment coupling control, thereby achieving optimized energy management.

[0006] In one embodiment, the multidimensional correlation modeling and pattern extraction of meteorological element features and time series features includes: Acquire at least a variety of historical meteorological data and historical load data, including ambient temperature, relative humidity, wind direction, wind speed, solar radiation, and atmospheric pressure; The Pearson correlation coefficients between various meteorological features and historical load data and other meteorological features are obtained, and meteorological features and time series features are selected as input parameters based on the correlation. Based on historical load data and various historical meteorological data, correlation screening and mapping relationships of time series and meteorological multivariate features are extracted. A deep neural network prediction model that integrates time series, meteorological multivariate features and user energy consumption habit demand data features is constructed to output a short-term load demand curve.

[0007] In one embodiment, the relevance filtering and mapping relationship extraction includes: When the correlation between a feature and a load is higher than a threshold and the redundancy between features is lower than a threshold, the feature is included in the input set. When multiple meteorological features meet the conditions at the same time, the feature with higher correlation to the load is selected to suppress redundancy and improve robustness.

[0008] In one embodiment, the data requirement features of the deep neural network prediction model include time series features and correlation features between shallow and deep meteorological features; The deep neural network prediction model includes a shallow feature extraction module, a deep feature extraction module, a time series feature extraction module, a cross-modal feature fusion module, and an output module. The shallow feature extraction module and the deep feature extraction module capture the shallow and deep dependencies of meteorological features on load changes; the time series feature extraction module extracts the periodic and trend information of time series features; the cross-modal feature fusion module performs multi-scale and cross-feature fusion according to different feature types; and the output module maps the fused features to load demand values.

[0009] In one embodiment, a multi-order RC building thermodynamic model is used, combining building thermal inertia, user thermal comfort requirements, and the prediction results to generate a dynamic feasible load domain characterizing personalized indoor state constraints. The implementation method is as follows: Obtain building heat transfer characteristic parameters, including building envelope style and heat transfer coefficient, window-to-wall ratio, and building thermal resistance; obtain building heat storage characteristic parameters, including building material heat capacity, building area, and heat storage coefficient. A multi-stage RC building thermodynamic model is constructed based on the parameters, and the model is calibrated using the prediction results under a set indoor temperature condition, so that it can characterize the thermal inertia characteristics of the building under a given indoor temperature constraint. Based on the prediction results and the multi-order RC building thermodynamic model, the thermal comfort demand data and load feasible region of building thermal inertia are characterized. The load feasible region includes the maximum range of user thermal comfort, the out-of-bounds penalty weight, and the comfort boundary distance.

[0010] In one embodiment, user-personalized settings are made, including setting the user's thermal comfort PMV range value and optimizing the system to operate with energy saving or cost saving as the optimization goal, which the user can choose according to their own needs; The user thermal comfort PMV range is set to ±1, ±0.5, and stable 0. The optimization goal is to achieve overall energy saving for all devices, i.e., minimum energy consumption, or overall cost saving, i.e., minimum cost optimization operation.

[0011] In one embodiment, the dual-layer collaborative optimization algorithm comprehensively considers building thermal inertia, personalized user thermal comfort needs, and energy-saving or cost-saving optimization regulation of multi-source equipment coupling control, including: Construct user energy consumption equipment scenarios, obtain equipment operating parameters, and build a real-time dynamic mapping relationship between equipment operating input parameters and load supply, energy consumption, and cost based on MLP; Users can set their own thermal comfort PMV value range and personalize their energy-saving or cost-saving goals; A Markov chain bi-level optimization model constrained by equipment operating boundaries and building thermal inertia and thermal comfort is established. By setting boundary penalty and reward models, an energy-saving and cost-saving strategy that considers building thermal inertia, user thermal comfort and multi-source equipment coupling control is generated and optimized.

[0012] In one embodiment, the two-layer optimization model includes: An upper-level reinforcement learning model is constructed to control indoor thermal inertia and thermal comfort optimization. The upper-level reinforcement learning model includes the load prediction results, reward function, reward function optimization objective, and reward function boundary conditions. Construct a lower-level heuristic or mathematical programming algorithm scheduling model to control the energy consumption optimization operation of the energy supply equipment of the residential integrated energy system. The scheduling model includes hourly load demand, objective function, and constraints on the objective function of the algorithm. The upper-layer reinforcement learning model interacts with the lower-layer heuristic or mathematical programming algorithm scheduling model to obtain the solution of the two-layer optimization model, thereby obtaining the indoor thermal comfort state and equipment energy consumption, and obtaining an optimized operation strategy.

[0013] In one embodiment of the optimization control method, the reward function in reinforcement learning is: Where, r boundary For boundary penalties, r cost r is the negative of the sum of electricity, gas, and other expenses for this hour; consumption It is the negative number of the sum of electricity, gas, and other energy consumption for this hour.

[0014] The optimization objective of the reward function in reinforcement learning is shown in the following formula: In the formula: Reward is the maximum daily cost or energy consumption including boundary penalties; t is each time step; N is the maximum time step; and reward is the maximum daily cost or energy consumption. t The reward at time step t is the maximum energy consumption or the negative of the cost, and the minimum cost or energy consumption, obtained while avoiding boundary penalties. The objective function of the heuristic algorithm is shown in the following formula: In the formula: E ele,t Let t be the power consumption of each device, and cost. ele,t Let E be the electricity price at time t. gas,t Let $t$ be the natural gas consumption at time $t$, and $cost$ be the cost of natural gas consumption at time $t$. gas For natural gas prices; The boundary conditions for the reward function in reinforcement learning are shown in the following equation: or Where: PMV is the predicted average perceived value index; PMV set The user sets a value, generally less than or equal to 1; ΔPMV is the change in the predicted average perceptual index at each time step; T min,set T represents the set minimum room temperature; T represents the room temperature; T max,set The maximum temperature set for the room; ΔT represents the degree of change in room temperature over time. The constraints described in the heuristic algorithm are as follows: In the formula: Q equip.min Minimum power load for the equipment; Q equip Q supplies power to the equipment. equip,max Q represents the maximum power load supplied to the equipment. demand This represents the load demand value.

[0015] In one embodiment, the two-layer collaborative optimization algorithm includes the following steps: Step 1: Read the user-defined thermal comfort constraints and optimization goals, and construct the upper layer of the two-layer optimization model to set the hourly thermal comfort PMV boundary and energy saving and cost saving goals through reinforcement learning; Step 2: Conduct two-layer optimization training based on user energy-consuming equipment, user optimization objectives, and load feasible domain. For the day-ahead user thermal comfort optimization method, the agent performs load demand control at each time step. The lower-layer control constraints of the two-layer optimization model achieve the minimum cost or minimum energy consumption reward per hour and are transmitted to the agent. Step 3: The agent continuously interacts with the environment, completing one round of operation, obtaining reward values ​​for each state and action change, and accumulating knowledge based on state, action, and reward. On this basis, the agent performs multiple iterative learning processes, accumulating a large amount of knowledge to learn the optimal strategy. Step 4: Obtain the optimal intelligent agent for optimized operation, obtain the optimal operation control load and equipment regulation under the optimization target, and obtain the optimization strategy under the optimization target.

[0016] Compared with the prior art, the beneficial effects of the present invention are: (1) Improved load forecasting accuracy and stability: This invention is based on a deep forecasting model that integrates multi-source spatiotemporal data and time series features. It improves the shortcomings of traditional statistical methods and simple machine learning in characterizing complex features and maintains higher forecasting accuracy and better inter-time stability under different energy consumption scenarios and extreme weather conditions.

[0017] (2) Improved regulation effect, robustness, and real-time performance: This invention proposes a two-layer optimization strategy that combines reinforcement learning with heuristic / mathematical programming algorithms. The two-layer model interactive learning not only solves the problems of poor real-time performance and insufficient robustness of traditional static optimization models, but also improves the ability of reinforcement learning to quickly locate effective policy intervals and reduce ineffective exploration, thereby improving solution efficiency, robustness, and real-time performance.

[0018] (3) Personalized user needs and improved thermal comfort: With the support of the aforementioned high-precision prediction and strong optimization capabilities, the flexible control of "virtual energy storage" is achieved by combining the building thermal inertia and thermal comfort model. It supports users to customize temperature / PMV comfort range and energy-saving and cost-saving modes, dynamically maintain or improve the physical comfort under constraints, and take into account energy consumption and cost targets to adapt to the personalized needs of different house types and equipment combinations. Attached Figure Description

[0019] Figure 1 A flowchart for energy consumption forecasting and optimization management methods for residential integrated energy systems.

[0020] Figure 2 This is a coupling diagram of multiple devices in a residential integrated energy system according to an example of the present invention.

[0021] Figure 3 This is a flowchart of the energy consumption prediction method for a residential integrated energy system in an example of the present invention.

[0022] Figure 4 This is a flowchart of the energy optimization management method for a residential integrated energy system in an example of the present invention.

[0023] Figure 5 This is a comparison chart of the energy prediction results of different control groups in the examples of this invention.

[0024] Figure 6 This is a comparison chart of energy optimization results for different control groups in the examples of this invention. Detailed Implementation

[0025] To more clearly illustrate the purpose, method, and advantages of this invention, further explanation and examples are provided in conjunction with the accompanying drawings.

[0026] This invention relates to an artificial intelligence-based method for predicting and intelligently managing energy consumption in residential integrated energy systems, such as... Figure 1 As shown. In some examples, the energy prediction method disclosed in this invention includes: S1, by performing multi-dimensional correlation modeling and pattern extraction on meteorological features and time series features, and then obtaining the prediction results of different load demands of users based on correlation screening and multi-scale feature fusion; S2 uses a multi-order RC building thermodynamic model, combining building thermal inertia, user thermal comfort requirements and the prediction results, to generate a dynamic feasible load domain that characterizes personalized indoor state constraints.

[0027] Furthermore, in S1 of this invention, the data to be acquired includes historical meteorological data and historical load data. The historical meteorological data includes ambient temperature, relative humidity, wind direction, wind speed, solar radiation, rainfall, and atmospheric pressure, while the historical load data includes at least user cooling, heating, and electricity load demands. These acquired data need to undergo data cleaning, data validity analysis, data alignment, and normalization. Subsequently, correlation calculations are performed, selecting meteorological features and time-series features as input parameters based on the correlation, i.e., selecting features with high correlation as input. Specifically, the correlation is represented by the Pearson correlation coefficient between each meteorological feature and load demand, as well as other meteorological features.

[0028] Based on historical load data and various historical meteorological data, the correlation between time series and meteorological multivariate features is screened and the mapping relationship is extracted. This includes: when the correlation between "feature-load" is higher than the threshold and the redundancy between "feature-feature" is lower than the threshold, it is included in the input set; when multiple meteorological features meet the conditions at the same time, those with higher correlation with load are selected first to suppress redundancy and improve robustness.

[0029] Finally, based on the selected input set and the corresponding mapping relationship, a deep neural network prediction model is constructed that integrates time series, meteorological multivariate features and user energy consumption habit demand data features. The output of this model is a short-term load demand curve, and the data demand features include time series features, deep and shallow meteorological features and correlation features.

[0030] Furthermore, in S2 of this invention, the data to be acquired includes building heat transfer and heat storage characteristic parameters, user energy consumption behavior patterns, meteorological data, and historical user energy demand data. The building heat transfer characteristic parameters mainly include building envelope style and heat transfer coefficient, window-to-wall ratio, and building thermal resistance, while the building heat storage characteristic parameters mainly include building material heat capacity, building area, and heat storage coefficient. Simultaneously, the relationship between equipment operating parameters and energy load output can be integrated to construct a mapping curve of the relationship between equipment environmental parameters, load demand, and energy consumption / cost. Equipment operating parameters include environmental parameters affecting equipment operation, equipment power generation parameters, load demand parameters, and equipment energy consumption / cost parameters.

[0031] In S2, a multi-order RC building thermodynamic model is used to calibrate the model when the indoor temperature is stable, incorporating the prediction results from S1. Subsequently, based on the prediction results and the building model, the thermal comfort demand data and load feasible region of the building's thermal inertia are characterized. The load feasible region includes constraint information such as the maximum range of user thermal comfort, the penalty weight for exceeding limits, and the comfort boundary distance. In this invention, the maximum range of user thermal comfort is set between -1 and 1, and the range of user thermal comfort variation is set to a PMV value change of less than 0.5 per hour.

[0032] Furthermore, user-personalized settings are also possible, including: setting the user's thermal comfort PMV range and optimizing the system to operate with energy saving or cost saving as the optimization goal, which users can choose according to their own needs. In some embodiments, the user's thermal comfort PMV range can be set to ±1, ±0.5, and stable 0 operation, while the optimization goal can be set to overall energy saving of each device (i.e., minimum energy consumption) or overall cost saving (i.e., minimum cost), optimizing operation.

[0033] Following on from the above, the intelligent control method of the present invention is as follows: S3 considers users' personalized needs for thermal comfort, energy saving, or cost saving, and constructs a two-layer collaborative optimization algorithm to build energy-saving and cost-saving strategies that take into account building thermal inertia, user thermal comfort, and multi-source equipment coupling control, thereby achieving optimized energy management.

[0034] For example, the two-layer collaborative optimization algorithm of the present invention comprehensively considers the building's thermal inertia and the personalized needs of users' thermal comfort, as well as the energy-saving or cost-saving optimization regulation of multi-source equipment coupling control, including: Construct user energy consumption equipment scenarios, and based on the mapping curves of equipment operating parameters and energy consumption and cost, use MLP to build real-time dynamic mapping relationships between equipment operating input parameters and load supply, energy consumption and cost. Users can set personalized settings such as PMV value range for thermal comfort and energy saving or cost saving targets. A Markov chain bi-level optimization model constrained by equipment operating boundaries and building thermal inertia and thermal comfort is established. By setting boundary penalty and reward models, a smart strategy for integrated energy saving and cost saving operation of residential buildings that considers building thermal inertia, user thermal comfort and multi-source equipment coupling control is generated and optimized.

[0035] In this embodiment, the constructed residential integrated energy system is as follows: Figure 2 The diagram shows models of air source heat pumps, gas-fired wall-hung boilers, fan coil units, and domestic water tanks. Based on equipment operating parameters, environmental parameters, load demand, and other data, a real-time dynamic mapping relationship is constructed between equipment operating input parameters and load supply, energy consumption, and costs, including: In the relational mapping, a mapping model of the relationship between ambient temperature, indoor temperature, load demand and power consumption of air source heat pump is constructed; a mapping curve of the relationship between load demand and gas volume of gas wall-hung boiler is constructed; a mapping curve of the relationship between load demand and energy consumption of fan coil unit is constructed; and a model of thermal resistance and electric heating of domestic water tank is constructed.

[0036] In this embodiment, the deep prediction model is a neural network model that integrates time series data, meteorological multivariate features, and user energy consumption habit demand data. The data demand features include time series features, shallow and deep meteorological features, and related features.

[0037] In this embodiment, the mapping relationship between meteorological multivariate characteristics and historical load data includes: Obtain the Pearson correlation coefficients between different meteorological characteristics, historical load data, and other meteorological characteristics; When the Pearson correlation coefficient between meteorological characteristics and historical load data is higher than 0.5, it is selected as the input parameter for meteorological multi-feature parameters. When the Pearson correlation coefficient between meteorological features is higher than 0.5 and the correlation coefficient between meteorological features and Pearson is higher than 0.5, meteorological features with a higher correlation to historical load data are selected as input parameters for meteorological multi-feature parameters.

[0038] In this embodiment, the meteorological multivariate characteristics consist of ambient temperature characteristics, relative humidity characteristics, wind direction characteristics, customs characteristics, and solar radiation intensity characteristics.

[0039] In this embodiment, the model mainly includes a shallow feature extraction module, a deep feature extraction module, a time series feature extraction module, a cross-modal feature fusion module, and an output module. Specifically: the shallow and deep feature extraction modules capture the shallow and deep dependencies of meteorological features on load changes; the time series feature extraction module extracts the periodicity and trend information of time series features; the cross-modal feature fusion module performs multi-scale, cross-feature fusion based on different feature types; and the output module maps the fused features to load demand values.

[0040] In this embodiment, the neural network model employs a multi-layer CNN to collect deep and shallow meteorological multivariate feature information and performs feature fusion to enhance the influence of meteorological features; a GRU neural network is used to collect energy load time series features. The above-mentioned different types of features are fused across multiple scales, and the fused features are mapped to load demand values ​​through an MLP network.

[0041] In this embodiment, the flowchart of the depth prediction model is as follows: Figure 3 As shown. The steps are as follows: Step 1: Based on historical data, obtain the Pearson correlation coefficients between environmental parameters and energy load, as well as the correlation coefficients between environmental parameters and the load. When the Pearson correlation coefficient between an environmental parameter and the load is higher than a threshold and the correlation coefficient between an environmental parameter and other environmental parameters is lower than a threshold, it is selected as the load forecasting feature.

[0042] Step 2: For environmental and load data from a limited number of past days, a CNN neural network is used to extract deep and shallow environmental features, and a GRU neural network is used to extract load time series features. The extracted environmental features and time series features are then fused across multiple scales and features.

[0043] Step 3: Start forward propagation on the training data, compare the results with the training data, and then start backpropagation to update and correct the parameters of different memory units in order to improve the accuracy of recent predictions.

[0044] Step 4: Input the environmental parameters required for prediction and the upstream load data into the trained model to perform day-ahead load prediction.

[0045] In this embodiment, the thermal comfort requirements data and load feasible domain of building thermal inertia are generated from building parameters and prediction results.

[0046] A multi-stage RC network model is constructed based on building parameters to dynamically regulate indoor temperature and load demand models, and to adjust according to comfort temperature or PMV range and comfort boundary.

[0047] The comfortable temperature range is 17~23℃, the PMV range is -1~1, and the comfort boundary is a temperature change of less than 2℃ per hour and a PMV change of less than 0.5 per hour.

[0048] In this embodiment, user personalization settings include user comfort temperature settings or user PMV range settings, which users can choose and select for cost-saving or energy-saving optimization according to their own needs.

[0049] Temperature settings can be set to 17~23 ℃, 19~21 ℃, or a stable 20 ℃. PMV range settings can be set to ±1, ±0.5, or a stable 0.

[0050] In this embodiment, the two-layer optimization model, based on the building's virtual energy storage load regulation and residential integrated energy system energy supply equipment energy consumption regulation strategies caused by thermal inertia and thermal comfort, includes: An upper-level reinforcement learning model is constructed to control indoor thermal inertia and thermal comfort optimization. The upper-level optimization scheduling model includes the load prediction results, the reward function, the reward function optimization objective, and the boundary conditions of the reward function through reinforcement learning. A lower-level heuristic or mathematical programming algorithm scheduling model is constructed to control the energy consumption optimization operation of the energy supply equipment of the residential integrated energy system. The scheduling model includes hourly load demand, objective function, and constraints on the objective function of the scheduling model. The upper-level reinforcement learning model interacts with the lower-level heuristic or mathematical programming algorithm scheduling model to obtain the solution of the two-level optimization model, thereby obtaining the indoor thermal comfort state and equipment energy consumption, and obtaining an optimized operation strategy.

[0051] In this embodiment, the reward function for reinforcement learning is specifically: Where, r boundary For boundary penalties, r cost r is the negative of the sum of electricity, gas, and other expenses for this hour; consumption It is the negative number of the sum of electricity, gas, and other energy consumption for this hour.

[0052] The optimization objective of the reward function in reinforcement learning is shown in the following formula: In the formula: Reward is the maximum daily cost or energy consumption including boundary penalties; t is each time step; N is the maximum time step; and reward is the maximum daily cost or energy consumption. t Let be the reward at time step t. The optimal result obtained, while avoiding boundary penalties, is the negative of the maximum energy consumption or cost, and the minimum cost or energy consumption.

[0053] The objective function of the scheduling model is shown in the following formula: In the formula: E ele,t Let t be the power consumption of each device, and cost.ele,t Let E be the electricity price at time t. gas,t Let $t$ be the natural gas consumption at time $t$, and $cost$ be the cost of natural gas consumption at time $t$. gas This refers to the price of natural gas.

[0054] The boundary conditions for the reward function in reinforcement learning are shown in the following equation: or Where: PMV is the predicted average perceived value index; PMV set The user sets a value, generally less than or equal to 1; ΔPMV is the change in the predicted average perceptual index at each time step; T min,set T represents the set minimum room temperature; T represents the room temperature; T max,set The maximum temperature set for the room; ΔT represents the degree of change in room temperature at each time step.

[0055] The constraints of the scheduling model are shown in the following formula: In the formula: Q equip.min Minimum power load for the equipment; Q equip Q supplies power to the equipment. equip,max Q represents the maximum power load supplied to the equipment. demand This represents the load demand value.

[0056] In this embodiment, the two-layer collaborative optimization algorithm flow is as follows: Figure 4 As shown, the steps are as follows: Step 1: Read the user-defined thermal comfort constraints and optimization goals, and construct the upper layer of the two-layer optimization model using reinforcement learning to set the hourly thermal comfort PMV boundary and energy saving and cost saving goals; Step 2: Conduct two-layer optimization training based on user energy-consuming equipment, user optimization objectives, and load feasible domain. For day-ahead thermal comfort or room temperature control methods, the agent performs load demand control at each time step. The lower-layer control constraints of the two-layer optimization model achieve the minimum cost or minimum energy consumption reward per hour, and transmit this information to the agent. Step 3: The agent continuously interacts with the environment, completing one round of operation, obtaining reward values ​​for each state and action change, and accumulating knowledge based on state, action, and reward. On this basis, the agent performs multiple iterative learning processes, accumulating a large amount of knowledge to learn the optimal strategy. Step 4: Obtain the optimal intelligent agent for optimized operation, obtain the optimal operation control load and equipment regulation under the optimization target, and obtain the optimization strategy under the optimization target.

[0057] Based on the above embodiments, this invention conducted comparative experiments on load prediction using a deep prediction model and experiments on cost-saving optimization. Two control groups were set up for the prediction model experiments, using CNN meteorological feature prediction and GRU time-series feature prediction respectively. Three control groups were set up for the cost-saving optimization experiments: standalone heat pump heating without considering thermal inertia, heat pump coupled with a wall-mounted boiler heating without considering thermal inertia, and standalone heat pump heating considering thermal inertia. These were used to compare and evaluate the load prediction and energy-saving or cost-saving optimization effects of this invention. The same parameters were selected for each control group. The experimental results are as follows: Figure 5 , Figure 6 As shown, compared to CNN and GRU load prediction alone, the deep prediction model improves the average accuracy of prediction for different cities by 2.52% and 0.61%, respectively. Compared to the three groups of single heat pump heating without considering thermal inertia, heat pump coupled wall-hung boiler heating without considering thermal inertia, and single heat pump heating considering thermal inertia, the optimization method proposed in this invention reduces the average cost for different cities by 12.47 yuan, 5.71 yuan, and 3.64 yuan, respectively.

[0058] In summary, this invention utilizes deep neural networks to comprehensively consider the influence of meteorological characteristics such as ambient temperature, relative humidity, and solar radiation, as well as time-series characteristics, to achieve multi-feature fusion load forecasting. Based on a multi-order RC model, it establishes a load feasible region by combining load forecasting with building thermal inertia and thermal comfort requirements. It employs reinforcement learning combined with mathematical programming / heuristic algorithms to comprehensively regulate residential energy consumption thermal inertia and personalized thermal comfort needs, as well as multi-source coupling optimization. This embodiment can balance users' personalized thermal comfort and energy needs, achieving optimized real-time performance and robustness under high-precision load supply, and can adapt to the personalized needs of different apartment types and equipment combinations.

[0059] This invention is not limited to the embodiments described above. The above description of specific embodiments is intended to illustrate and explain the technical solutions of this invention. The specific embodiments described above are merely illustrative and not restrictive. Without departing from the spirit and scope of the claims, those skilled in the art can make many specific modifications based on the teachings of this invention, and these modifications all fall within the scope of protection of this invention.

Claims

1. A method for predicting and intelligently managing energy consumption in a residential integrated energy system based on artificial intelligence, characterized in that, Includes the following steps: By performing multidimensional correlation modeling and pattern extraction of meteorological and time series features, and based on correlation screening and multi-scale feature fusion, the predicted results of different user load demands are obtained. A multi-order RC building thermodynamic model is adopted, and the building thermal inertia, user thermal comfort requirements and the prediction results are combined to generate a dynamic feasible load domain that characterizes the individual indoor state constraints. Considering users' personalized needs for thermal comfort, energy saving, and cost reduction, a two-layer collaborative optimization algorithm is constructed to develop energy-saving and cost-saving strategies that take into account building thermal inertia, user thermal comfort, and multi-source equipment coupling control, thereby achieving optimized energy management.

2. The method for energy consumption prediction and intelligent management of a residential integrated energy system based on artificial intelligence as described in claim 1, characterized in that, The multidimensional correlation modeling and pattern extraction of meteorological element characteristics and time series characteristics includes: Acquire at least a variety of historical meteorological data and historical load data, including ambient temperature, relative humidity, wind direction, wind speed, solar radiation, and atmospheric pressure; The Pearson correlation coefficients between various meteorological features and historical load data and other meteorological features are obtained, and meteorological features and time series features are selected as input parameters based on the correlation. Based on historical load data and various historical meteorological data, correlation screening and mapping relationships of time series and meteorological multivariate features are extracted. A deep neural network prediction model that integrates time series, meteorological multivariate features and user energy consumption habit demand data features is constructed to output a short-term load demand curve.

3. The method for predicting and intelligently managing energy consumption in a residential integrated energy system based on artificial intelligence as described in claim 2, characterized in that, The correlation filtering and mapping relationship extraction include: When the correlation between "feature and load" is higher than the threshold and the redundancy between "feature and feature" is lower than the threshold, it is included in the input set; when multiple meteorological features meet the conditions at the same time, the one with higher correlation with load is selected to suppress redundancy and improve robustness.

4. The method for energy consumption prediction and intelligent management of a residential integrated energy system based on artificial intelligence as described in claim 2, characterized in that, The data requirements of the deep neural network prediction model include time series features, shallow and deep meteorological features, and related features. The deep neural network prediction model includes a shallow feature extraction module, a deep feature extraction module, a time series feature extraction module, a cross-modal feature fusion module, and an output module. The shallow feature extraction module and the deep feature extraction module capture the shallow and deep dependencies of meteorological features on load changes; the time series feature extraction module extracts the periodic and trend information of time series features; and the cross-modal feature fusion module performs multi-scale and cross-feature fusion according to different feature types. The output module maps the fused features to load demand values.

5. The method for predicting and intelligently managing energy consumption in a residential integrated energy system based on artificial intelligence as described in claim 1, characterized in that, The method employs a multi-order RC building thermodynamic model, combining building thermal inertia, user thermal comfort requirements, and the predicted results to generate a dynamic feasible load domain characterizing personalized indoor state constraints. The implementation method is as follows: Obtain building heat transfer characteristic parameters, including building envelope style and heat transfer coefficient, window-to-wall ratio, and building thermal resistance; obtain building heat storage characteristic parameters, including building material heat capacity, building area, and heat storage coefficient. A multi-stage RC building thermodynamic model is constructed based on the parameters, and the model is calibrated using the prediction results under a set indoor temperature condition, so that it can characterize the thermal inertia characteristics of the building under a given indoor temperature constraint. Based on the prediction results and the multi-order RC building thermodynamic model, the thermal comfort demand data and load feasible region of building thermal inertia are characterized. The load feasible region includes the maximum range of user thermal comfort, the out-of-bounds penalty weight, and the comfort boundary distance.

6. The method for energy consumption prediction and intelligent management of a residential integrated energy system based on artificial intelligence according to claim 5, characterized in that, Personalized settings can be made for users, including setting the PMV range value for thermal comfort and optimizing the system to operate with energy saving or cost saving as the optimization goal. Users can choose according to their own needs. The user thermal comfort PMV range is set to ±1, ±0.5, and stable 0. The optimization goal is to achieve overall energy saving for all devices, i.e., minimum energy consumption, or overall cost saving, i.e., minimum cost optimization operation.

7. The method for energy consumption prediction and intelligent management of a residential integrated energy system based on artificial intelligence as described in claim 1, characterized in that, The aforementioned dual-layer collaborative optimization algorithm comprehensively considers building thermal inertia, personalized user thermal comfort needs, and energy-saving or cost-saving optimization regulation of multi-source equipment coupling control, including: Construct user energy consumption equipment scenarios, obtain equipment operating parameters, and build a real-time dynamic mapping relationship between equipment operating input parameters and load supply, energy consumption, and cost based on MLP; Users can set their own thermal comfort PMV value range and personalize their energy-saving or cost-saving goals; A Markov chain bi-level optimization model constrained by equipment operating boundaries and building thermal inertia and thermal comfort is established. By setting boundary penalty and reward models, an energy-saving and cost-saving strategy that considers building thermal inertia, user thermal comfort and multi-source equipment coupling control is generated and optimized.

8. The method for predicting and intelligently managing energy consumption in a residential integrated energy system based on artificial intelligence as described in claim 7, characterized in that, The two-layer optimization model includes: An upper-level reinforcement learning model is constructed to control indoor thermal inertia and thermal comfort optimization. The upper-level reinforcement learning model includes the load prediction results, reward function, reward function optimization objective, and reward function boundary conditions. Construct a lower-level heuristic or mathematical programming algorithm scheduling model to control the energy consumption optimization operation of the energy supply equipment of the residential integrated energy system. The scheduling model includes hourly load demand, objective function, and constraints on the objective function of the algorithm. The upper-layer reinforcement learning model interacts with the lower-layer heuristic or mathematical programming algorithm scheduling model to obtain the solution of the two-layer optimization model, thereby obtaining the indoor thermal comfort state and equipment energy consumption, and obtaining an optimized operation strategy.

9. The method for predicting and intelligently managing energy consumption in a residential integrated energy system based on artificial intelligence as described in claim 8, characterized in that, In the optimized control method, the reward function in reinforcement learning is: Where, r boundary For boundary penalties, r cost r is the negative of the sum of electricity, gas, and other expenses for this hour; consumption It is the negative number of the sum of electricity, gas, and other energy consumption for this hour. The optimization objective of the reward function in reinforcement learning is shown in the following formula: In the formula: Reward is the maximum daily cost or energy consumption including boundary penalties; t is each time step; N is the maximum time step; and reward is the maximum daily cost or energy consumption. t The reward at time step t is the maximum energy consumption or the negative of the cost, and the minimum cost or energy consumption, obtained while avoiding boundary penalties. The objective function of the heuristic algorithm is shown in the following formula: In the formula: E ele,t Let t be the power consumption of each device, and cost. ele,t Let E be the electricity price at time t. gas,t Let $t$ be the natural gas consumption at time $t$, and $cost$ be the cost of natural gas consumption at time $t$. gas For natural gas prices; The boundary conditions for the reward function in reinforcement learning are shown in the following equation: or Where: PMV is the predicted average perceived value index; PMV set The user sets a value, generally less than or equal to 1; ΔPMV is the change in the predicted average perceptual index at each time step; T min,set T represents the set minimum room temperature; T represents the room temperature; T max,set The maximum temperature set for the room; ΔT represents the degree of change in room temperature over time. The constraints described in the heuristic algorithm are as follows: In the formula: Q equip.min Minimum power load for the equipment; Q equip Q supplies power to the equipment. equip,max Q represents the maximum power load supplied to the equipment. demand This represents the load demand value.

10. The method for energy consumption prediction and intelligent management of a residential integrated energy system based on artificial intelligence according to claim 7, characterized in that, The two-layer collaborative optimization algorithm includes the following steps: Step 1: Read the user-defined thermal comfort constraints and optimization goals, and construct the upper layer of the two-layer optimization model to set the hourly thermal comfort PMV boundary and energy saving and cost saving goals through reinforcement learning; Step 2: Conduct two-layer optimization training based on user energy-consuming equipment, user optimization objectives, and load feasible domain; for the day-ahead user thermal comfort optimization method, the agent performs load demand control at each time step, and the lower-level control constraints of the two-layer optimization model achieve the minimum cost or minimum energy consumption reward per hour, and transmit it to the agent; Step 3: The agent continuously interacts with the environment to complete one round of operation, obtains the reward value for each change in state and action, and accumulates knowledge based on state, action, and reward; on this basis, the agent performs multiple iterative learning to accumulate knowledge in order to learn the optimal strategy. Step 4: Obtain the optimal intelligent agent for optimized operation, obtain the optimal operation control load and equipment regulation under the optimization target, and obtain the optimization strategy under the optimization target.

Images