Intelligent building multi-energy flow joint optimization scheduling method and system based on artificial intelligence

Abstract

The application provides an artificial intelligence-based intelligent building multi-energy flow joint optimization scheduling method and system, relates to the technical field of data processing, and comprises the following steps: processing historical multi-source data of an intelligent building, constructing a differentiated training data set, analyzing the differentiated training data set to obtain a target hourly prediction curve of the intelligent building in a target scheduling period, constructing a double-layer optimization scheduling architecture with an energy operator as a leader and a building user as a follower to perform optimization scheduling analysis on a photovoltaic power generation power curve, a building electrical load curve and a cold and heat load curve, and obtaining an optimal scheduling scheme to perform optimization scheduling on the intelligent building. The application solves the technical problem that the existing multi-energy flow scheduling lacks a collaborative optimization mechanism, leading to high energy consumption and difficulty in balancing comfort. Through multi-source data intelligent prediction and double-layer optimization scheduling, the intelligent building multi-energy flow collaborative optimization is realized, and the technical effects of improving energy utilization efficiency and indoor comfort are achieved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and system for joint optimization scheduling of multi-energy flow in smart buildings based on artificial intelligence. Background Technology

[0002] As building energy consumption accounts for an increasingly larger share of total urban energy consumption, smart buildings, as integrated energy systems combining photovoltaic power generation, energy storage devices, heating and cooling systems, and controllable loads, are gradually becoming an important means to improve energy efficiency and achieve carbon emission reduction. However, existing building energy management methods mostly target the scheduling of single energy types, lacking a holistic coordination and optimization mechanism for multiple energy flows such as electricity, cooling, and heating, resulting in high system energy consumption and low energy efficiency. Furthermore, existing methods struggle to achieve coordinated control of indoor temperature and humidity, lighting environment, and user comfort, typically treating comfort as a rigid constraint and failing to accurately reflect the continuous changes in human perception of the thermal environment and lighting, thus impacting user comfort while pursuing energy conservation and cost reduction.

[0003] Existing technologies suffer from a lack of collaborative optimization mechanisms in multi-energy flow scheduling, resulting in high energy consumption and difficulty in achieving both energy efficiency and comfort. Summary of the Invention

[0004] The purpose of this application is to provide a method and system for joint optimization scheduling of multi-energy flow in smart buildings based on artificial intelligence, in order to solve the technical problem that the existing technology lacks a collaborative optimization mechanism for multi-energy flow scheduling, resulting in high energy consumption and difficulty in balancing comfort.

[0005] In view of the above problems, this application provides a method and system for joint optimization scheduling of multi-energy flow in smart buildings based on artificial intelligence.

[0006] The first aspect of this application provides an AI-based method for joint optimization scheduling of multi-energy flows in smart buildings. This method includes: cleaning and normalizing historical multi-source data of the smart building to construct a differentiated training dataset; analyzing the differentiated training dataset to obtain the target hourly prediction curve of the smart building in a target scheduling period, wherein the target hourly prediction curve includes a photovoltaic power generation curve, a building electrical load curve, and a heating / cooling load curve; constructing a two-layer optimization scheduling architecture with energy operators as leaders and building users as followers; performing optimization scheduling analysis on the photovoltaic power generation curve, the building electrical load curve, and the heating / cooling load curve through the two-layer optimization scheduling architecture to obtain an optimal scheduling scheme; and executing optimized scheduling of the smart building according to the optimal scheduling scheme in the target scheduling period.

[0007] Optionally, a first historical data group is extracted from the historical multi-source data, and outlier removal and unit unification are performed on the first historical data group to obtain a first standardized dataset; the first standardized dataset is divided based on the four-season division strategy to obtain historical division results; the spring dataset, summer dataset, autumn dataset, and winter dataset in the historical division results are combined to form the differentiated training dataset; wherein, the first historical data group includes first historical meteorological data, first historical building internal environment data, first historical multi-energy load data, and first historical distributed power output data.

[0008] Optionally, the backpropagation neural network is initialized with first initial weights and first initial thresholds; using the first initial weights and first initial thresholds as optimization objects, first target weights and first target thresholds are obtained through particle swarm optimization; the first target weights and first target thresholds are substituted into the backpropagation neural network to construct a first prediction network, a second prediction network, and a third prediction network; the differentiated training dataset is input into the first prediction network, the second prediction network, and the third prediction network, and the photovoltaic power generation curve, the building electrical load curve, and the heating and cooling load curve are obtained through forward propagation calculation.

[0009] Optionally, the solar radiation intensity, ambient temperature, and photovoltaic power generation at the same time the day before the prediction date are extracted as the first input parameters; the solar radiation intensity and ambient temperature at the same time the prediction date are obtained as the second input parameters; the first input parameters and the second input parameters are concatenated and input into the first prediction network, and the photovoltaic power generation curve is obtained through forward propagation calculation.

[0010] Optionally, with maximizing the difference between the energy operator's energy sales revenue and carbon emission costs as the first optimization objective, and using the first electricity purchase price and the first electricity sales price set for the building users as the first decision variables, an upper-level pricing optimization model is generated; with minimizing the building users' energy costs, maximizing overall satisfaction, and minimizing carbon emissions as the second optimization objective, a lower-level energy consumption optimization model is generated; wherein, the overall satisfaction includes thermal comfort satisfaction and lighting environment satisfaction; and the two-layer optimized scheduling architecture is constructed based on the upper-level pricing optimization model and the lower-level energy consumption optimization model.

[0011] Optionally, the indoor temperature and humidity of the smart building are obtained to obtain a first environmental parameter; the first environmental parameter is substituted into the predicted average vote count index to obtain a first voting index; the first environmental parameter is substituted into the predicted dissatisfaction percentage index to obtain a first dissatisfaction index; and the thermal comfort satisfaction is calculated based on the first voting index and the first dissatisfaction index.

[0012] Optionally, the total ambient illuminance of the smart building is obtained; the total ambient illuminance is substituted into a Gaussian distribution function to calculate the lighting comfort index; and the lighting comfort index is used as the lighting environment satisfaction.

[0013] Optionally, the rate of change of indoor temperature within the allowable range is extracted to construct virtual charging and discharging power; the upper and lower limits of adjustable indoor temperature are extracted to construct virtual capacity; and the indoor temperature comfort constraint is generated based on the virtual charging and discharging power and the virtual capacity.

[0014] Optionally, the first electricity purchase price and the first electricity sales price are updated using a genetic algorithm to obtain a first electricity price strategy; based on the first electricity price strategy, a mathematical programming solver is called to optimize the energy consumption strategy of the building user to obtain a first energy consumption strategy; the first electricity price strategy and the first energy consumption strategy are iterated alternately, and the change in the objective function value between two consecutive iterations is extracted; when the change in the objective function value is less than a preset threshold, it is determined that a Nash equilibrium has been reached, and the optimal scheduling scheme is generated; wherein, after generating the optimal scheduling scheme, the process further includes: extracting the equipment start-up and shutdown plan, energy storage charging and discharging plan, grid interaction power plan, flexible load response scheme, and indoor temperature setpoint from the optimal scheduling scheme; based on the thermal inertia and cold inertia characteristics of the building envelope, a virtual charging and discharging power model and a virtual capacity model of the building's virtual energy storage device are established; the indoor temperature setpoint is substituted into the virtual charging and discharging power model and the virtual capacity model to verify whether the indoor temperature setpoint meets the indoor temperature comfort constraint; if it does, the equipment start-up and shutdown plan, the energy storage charging and discharging plan, the grid interaction power plan, and the flexible load response scheme are output as the optimal scheduling scheme.

[0015] A second aspect of this application provides an AI-based smart building multi-energy flow joint optimization scheduling system, comprising: a dataset construction module for cleaning and normalizing historical multi-source data of the smart building to construct a differentiated training dataset; a prediction curve acquisition module for analyzing the differentiated training dataset to obtain the target hourly prediction curve of the smart building in the target scheduling period, wherein the target hourly prediction curve includes a photovoltaic power generation curve, a building electrical load curve, and a heating / cooling load curve; a scheduling architecture construction module for constructing a two-layer optimized scheduling architecture with energy operators as leaders and building users as followers; an optimized scheduling analysis module for performing optimized scheduling analysis on the photovoltaic power generation curve, the building electrical load curve, and the heating / cooling load curve through the two-layer optimized scheduling architecture to obtain an optimal scheduling scheme; and an optimized scheduling execution module for performing optimized scheduling execution on the smart building according to the optimal scheduling scheme in the target scheduling period.

[0016] One or more technical solutions provided in this application have at least the following technical effects or advantages: The method provided in this application cleans and normalizes historical multi-source data of smart buildings to construct a differentiated training dataset; analyzes the differentiated training dataset to obtain the target hourly prediction curve of the smart building in the target scheduling period, wherein the target hourly prediction curve includes photovoltaic power generation curve, building electrical load curve, and heating and cooling load curve; constructs a two-layer optimized scheduling architecture with energy operators as leaders and building users as followers; performs optimized scheduling analysis on the photovoltaic power generation curve, building electrical load curve, and heating and cooling load curve through the two-layer optimized scheduling architecture to obtain the optimal scheduling scheme; and executes optimized scheduling of the smart building according to the optimal scheduling scheme in the target scheduling period. This achieves the technical effect of realizing multi-energy flow collaborative optimization of smart buildings through intelligent prediction of multi-source data and two-layer optimized scheduling, thereby improving energy utilization efficiency and indoor comfort.

[0017] The above description is merely an overview of the technical solution of this application. To better understand the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 A flowchart illustrating the AI-based smart building multi-energy flow joint optimization scheduling method provided in this application.

[0020] Figure 2 The photovoltaic power generation curve of the target scheduling period in the AI-based smart building multi-energy flow joint optimization scheduling method provided in this application.

[0021] Figure 3 A schematic diagram of the structure of the AI-based smart building multi-energy flow joint optimization scheduling system provided in this application.

[0022] Figure labeling: Dataset construction module 11, Prediction curve acquisition module 12, Scheduling architecture construction module 13, Optimized scheduling analysis module 14, Optimized scheduling execution module 15. Detailed Implementation

[0023] This application provides an artificial intelligence-based method and system for the joint optimization and scheduling of multi-energy flows in smart buildings. It addresses the technical problem of existing technologies lacking a collaborative optimization mechanism in multi-energy flow scheduling, leading to high energy consumption and difficulty in balancing comfort. The method achieves the technical effect of improving energy efficiency and indoor comfort by using intelligent prediction based on multi-source data and two-layer optimization scheduling to realize collaborative optimization of multi-energy flows in smart buildings.

[0024] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. It should also be noted that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, not all of them.

[0025] Example 1, as Figure 1 As shown, this application provides a multi-energy flow joint optimization scheduling method for smart buildings based on artificial intelligence. The method includes: Historical multi-source data of smart buildings are cleaned and normalized to construct a differentiated training dataset.

[0026] Furthermore, the historical multi-source data of the smart building is cleaned and normalized to construct a differentiated training dataset, including: extracting a first historical data group from the historical multi-source data, and removing outliers and unifying the dimensions of the first historical data group to obtain a first standardized dataset; dividing the first standardized dataset based on a four-season division strategy to obtain historical division results; and combining the spring dataset, summer dataset, autumn dataset, and winter dataset from the historical division results to form the differentiated training dataset; wherein, the first historical data group includes first historical meteorological data, first historical building internal environment data, first historical multi-energy load data, and first historical distributed power output data.

[0027] Specifically, historical multi-source data covering at least four consecutive natural quarters is acquired from the data acquisition and monitoring control system deployed in the smart building. This historical multi-source data includes time-series data collected by terminal devices such as meteorological monitoring stations, indoor environmental monitoring systems, building energy management systems, and distributed power inverters and energy storage management systems. A first historical data group is extracted from this historical multi-source data. This first historical data group includes first historical meteorological data, first historical building internal environmental data, first historical multi-energy load data, and first historical distributed power output data. The first historical meteorological data includes parameters such as outdoor temperature, humidity, and solar radiation intensity. The first historical building internal environmental data includes indoor air temperature, humidity, illuminance, and CO2 concentration. The first historical multi-energy load data includes the building's historical total electrical power, cooling load, and heating load. The first historical distributed power output data mainly consists of photovoltaic, wind power, or energy storage discharge power.

[0028] For each parameter class, the upper and lower quartiles and interquartile ranges of its historical sequence are calculated using the interquartile range method. Samples with values ​​less than the lower quartile minus 1.5 times the interquartile range or greater than the upper quartile plus 1.5 times the interquartile range are marked as outliers and removed. Missing positions are filled using linear interpolation or forward imputation. Then, the min-max normalization method is used to map the data range of each parameter class to the interval [0,1]. The minimum and maximum values ​​calculated are the minimum and maximum values ​​of each parameter class in the historical sequence, respectively. By unifying the units of measurement, the influence of different physical units on the numerical scale is eliminated, resulting in the first standardized dataset.

[0029] The first standardized dataset is partitioned based on a four-season division strategy, defined according to meteorological or calendar quarters, such as spring (March to May), summer (June to August), autumn (September to November), and winter (December to February of the following year). Each data point in the first standardized dataset is assigned to its corresponding seasonal subset based on its timestamp, resulting in historical partitioning results. These historical partitioning results include spring, summer, autumn, and winter datasets. The seasonal datasets from these historical partitioning results are then combined to form a differentiated training dataset, which preserves the differences in data distribution under different seasonal conditions.

[0030] By constructing differentiated training datasets through cleaning, normalization, and seasonal differentiation, we can enhance the generalization ability of artificial intelligence under different environmental conditions, avoid seasonal bias caused by single training data, improve the accuracy of photovoltaic power generation, building load, and heating and cooling load prediction, and provide a reliable basis for optimized scheduling.

[0031] The differential training dataset is analyzed to obtain the target hourly prediction curve of the smart building in the target scheduling cycle, wherein the target hourly prediction curve includes the photovoltaic power generation curve, the building electrical load curve, and the cooling and heating load curve.

[0032] Furthermore, the analysis of the differentiated training dataset to obtain the target hourly prediction curve of the smart building during the target scheduling cycle includes: initializing the first initial weights and the first initial threshold of the backpropagation neural network; using the first initial weights and the first initial threshold as optimization objects, obtaining the first target weights and the first target threshold through particle swarm optimization; substituting the first target weights and the first target threshold into the backpropagation neural network to construct the first prediction network, the second prediction network, and the third prediction network respectively; inputting the differentiated training dataset into the first prediction network, the second prediction network, and the third prediction network respectively, and obtaining the photovoltaic power generation curve, the building electrical load curve, and the heating and cooling load curve through forward propagation calculation.

[0033] Furthermore, the differentiated training dataset is input into the first prediction network, and the photovoltaic power generation curve is obtained through forward propagation calculation, including: extracting the solar radiation intensity, ambient temperature, and photovoltaic power generation at the same time the day before the prediction date as the first input parameters; obtaining the solar radiation intensity and ambient temperature at the same time on the prediction date as the second input parameters; concatenating the first input parameters and the second input parameters and inputting them into the first prediction network, and obtaining the photovoltaic power generation curve through forward propagation calculation.

[0034] Specifically, artificial intelligence predictive analysis is performed on the differentiated training dataset to obtain the target hourly prediction curve of the smart building within the target scheduling period. The target scheduling period is the next 24 hours, but can also be set to 48 hours or 72 hours according to actual scheduling needs. The target hourly prediction curve refers to the future time-series prediction result formed at fixed time intervals, such as 1 hour.

[0035] In the analysis of the differential training dataset, a three-layer backpropagation neural network is first constructed, consisting of an input layer, a hidden layer, and an output layer. The input layer is responsible for receiving historical data sets, and the hidden layer is responsible for extracting nonlinear feature relationships. The number of nodes can be determined using an empirical formula. The algorithm is defined as follows: H is the number of hidden layer nodes, m is the number of input nodes, n is the number of output nodes, and a is an adjustment constant between 1 and 10. Each hidden layer node performs a weighted summation of the received input vector with its corresponding weights, adds a neuron threshold, and then maps the weighted summation result to the hidden layer output through a Sigmoid activation function. The output layer receives the hidden layer output and performs a weighted summation using the weights and thresholds from the hidden layer to the output layer. The Purelin function is used to directly output the predicted value at the target time, such as photovoltaic power generation, building electrical load, or heating / cooling load. The number of output layer nodes corresponds to the number of predicted targets. A backpropagation neural network is a feedforward artificial neural network that adjusts weights and thresholds through an error backpropagation mechanism. Its training process includes two stages: forward propagation to calculate the output and backpropagation to update the parameters, to achieve a nonlinear mapping relationship between input features and output targets. The weights include the connection weights from the input layer to the hidden layer and the connection weights from the hidden layer to the output layer, and the thresholds include the hidden layer neuron thresholds and the output layer neuron thresholds. The Xavier uniform initialization method is used to randomly initialize all weights and thresholds in the backpropagation neural network to obtain the first initial weights and the first initial threshold, ensuring that the parameters of the backpropagation neural network have a reasonable distribution range in the initial stage and avoiding excessively large or small gradients.

[0036] The first initial weights and the first initial thresholds are combined according to the encoding rules to form a particle vector, which is then used as the optimization target of the particle swarm optimization algorithm. An initial particle swarm is constructed based on the particle vectors, where each particle corresponds to a complete set of network weights and threshold parameters. For example, the population size is set to 30, the maximum number of iterations is set to 100, the inertia weight adopts a linear decreasing strategy, decreasing linearly from 0.9 to 0.4, and the acceleration factor is set to 2.0. During the optimization process, the parameter vector corresponding to each particle is decoded into the actual weights and thresholds of the backpropagation neural network and loaded into the backpropagation neural network. A differentiated training dataset is used to perform forward propagation calculations on the backpropagation neural network to obtain the predicted output. The mean squared error between the predicted output and the actual sample labels is calculated, and this mean squared error is used as the fitness value of the particle. The smaller the fitness value, the higher the prediction accuracy of the network parameter combination corresponding to the current particle. In each iteration, the individual historical best position of each particle and the global best position of the entire particle swarm are recorded. Based on the velocity update formula and position update formula of the particle swarm optimization algorithm, the flight speed and search position of each particle are dynamically adjusted so that the particles gradually converge toward a better solution. When the maximum number of iterations is reached or the change in global best fitness over several consecutive iterations is less than the preset convergence threshold of 0.01, the particle swarm optimization process is considered to have ended, and the weight and threshold corresponding to the global best particle are used as the first target weight and the first target threshold.

[0037] The first objective weights and the first objective threshold, which are optimized separately, are substituted into three backpropagation neural networks with the same structure but trained independently to construct a first prediction network, a second prediction network, and a third prediction network. The first prediction network is used to predict photovoltaic power generation, the second prediction network is used to predict building electrical load, and the third prediction network is used to predict heating and cooling load.

[0038] The differentiated training dataset is then input into the first, second, and third prediction networks for analysis. Historical operational information for the same time the day before the prediction date, including solar radiation intensity, ambient temperature, and photovoltaic power generation, is extracted from the differentiated training dataset as the first input parameter. Simultaneously, meteorological forecast data for the same time on the prediction date, including solar radiation intensity and ambient temperature, is obtained and normalized as the second input parameter. The first and second input parameters are concatenated in chronological order to form a 5-dimensional network input vector, which is then input into the first prediction network. The first prediction network calculates the photovoltaic power prediction value for the target time through forward propagation. Then, for each time point within the target scheduling period, the above input extraction and forward propagation calculation are repeated. The prediction outputs for each time point are arranged in chronological order to obtain the photovoltaic power generation curve of the smart building within the target scheduling period. For example, by calculating hourly for 24 consecutive scheduling periods, a photovoltaic power generation curve for the next 24 hours is formed.

[0039] Similarly, the second prediction network takes historical electrical load, outdoor temperature, humidity, and building interior environmental data as inputs and outputs future hourly building electrical load curves, while the third prediction network takes historical cooling load, heating load, outdoor temperature, humidity, and time attributes as inputs and outputs future hourly cooling and heating load curves.

[0040] For example, if the target scheduling date is April 16, 2025, and the time resolution is 1 hour, when predicting the photovoltaic power generation at 14:00 on that day, the first step is to prepare the input data: extract the normalized solar radiation intensity (0.75), ambient temperature (0.85), and measured photovoltaic inverter power generation (0.603) at 14:00 on the day before the prediction date (April 15) from the differentiated training dataset, which constitute the first input parameters. Simultaneously, obtain the predicted solar radiation intensity (798 W / m²) at 14:00 on the prediction date (April 16) from the numerical weather prediction interface. 2Given a predicted ambient temperature of 33.8℃, and based on historical extreme values ​​saved during data preprocessing, the radiation intensity range is 0~1100W / m², the temperature range is -10~42℃, and the photovoltaic power range is 0~120kW. Normalization is performed on these parameters, mapping them to the [0,1] interval. Specifically, the predicted daily radiation intensity is normalized to 0.725, and the predicted daily temperature is normalized to 0.842, forming the second input parameter. The first input parameter and the second input parameter are concatenated to form a 5-dimensional input vector [0.750, 0.850, 0.603, 0.725, 0.842]. T The input to the first prediction network is used for forward propagation calculation: the input layer of the first prediction network is passed to the hidden layer, and linearly combined with the first target weight matrix and the first target threshold vector obtained from joint training of particle swarm optimization and backpropagation. After activation, the hidden layer output vector is obtained, and then the output layer outputs a normalized prediction value of 0.587. After inverse normalization, it is converted into an actual photovoltaic power of 70.4kW. The inverse normalization formula is: P actual =P norm ×(P max -P min )+P min , where P norm =0.587 is the normalized predicted value, P max P min The maximum power values ​​are 120 respectively. kW and minimum power value 0 kW P actual This represents the actual photovoltaic power. Finally, a prediction curve is generated: For each of the 24 times from 00:00 to 23:00 on April 16th, the entire process of parameter extraction, normalization, forward propagation, and denormalization is repeated. The predicted values ​​for each time point are arranged in chronological order, forming: [0.0,0.0,0.0,0.0,2.1,15.3,38.7,58.2,72.5,79.8,82.1,78.6,71.3,70.4,63.8,52.4,37.9,21.5,8.2,1.3,0.0,0.0,0.0,0.0]kW. This corresponds to the complete photovoltaic power generation curve for that day, as shown below. Figure 2 As shown, the photovoltaic power generation curve represents the predicted photovoltaic power generation values ​​for 24 hours from 00:00 to 23:00 on April 16, 2025 (the target scheduling day). The horizontal axis represents the time of the prediction day, and the vertical axis represents the photovoltaic power generation. 00:00-03:00 and 20:00-23:00 are the nighttime zero-output periods with no solar radiation, and the photovoltaic power generation is always 0kW. 04:00-10:00 is the rapid increase period, and 11:00-19:00 is the slow decrease period.

[0041] By establishing a predictive network for photovoltaic power output, building electrical load, and heating and cooling load through particle swarm optimization backpropagation neural network, subsequent optimization scheduling no longer relies on static empirical values, but makes decisions based on future energy supply and demand trends, thereby improving the foresight and accuracy of scheduling schemes.

[0042] Construct a two-tier optimized scheduling architecture with energy operators as leaders and building users as followers.

[0043] Furthermore, a two-tiered optimized scheduling architecture is constructed, with energy operators as leaders and building users as followers. This includes: maximizing the difference between the energy operator's energy sales revenue and carbon emission costs as the first optimization objective, and using the first electricity purchase price and the first electricity sales price set for the building users as the first decision variables to generate an upper-level pricing optimization model; minimizing the building users' energy costs, maximizing overall satisfaction, and minimizing carbon emissions as the second optimization objective to generate a lower-level energy consumption optimization model; wherein, the overall satisfaction includes thermal comfort satisfaction and lighting environment satisfaction; and constructing the two-tiered optimized scheduling architecture based on the upper-level pricing optimization model and the lower-level energy consumption optimization model.

[0044] Furthermore, the constraints of the lower-level energy optimization model include indoor temperature comfort constraints, including: extracting the rate of change of indoor temperature within the allowable range and constructing virtual charging and discharging power; extracting the upper and lower limits of adjustable indoor temperature and constructing virtual capacity; and generating the indoor temperature comfort constraints based on the virtual charging and discharging power and the virtual capacity.

[0045] Specifically, energy operators, as leaders, prioritize maximizing the difference between energy sales revenue and carbon emission costs. Energy sales revenue is obtained by summing the products of electricity purchased by building users and the actual electricity price, i.e.: , where R sale For revenue from energy sales, P sale (t) represents the unit price (yuan / kWh) of electricity sold to building users at time t. sale (t) represents the actual electricity purchased by the building user at time t (kWh), and T is the total number of time periods in the dispatch cycle. The carbon emission cost Cco2 is calculated using the carbon emission factor α (kgCO2 / kWh) corresponding to the building user's electricity consumption and the unit carbon emission cost β (yuan / kgCO2). And the initial electricity purchase price P set for building users. buy With the first electricity sales price P sell As the primary decision variable, the initial electricity purchase price P buy It is the electricity purchase price set by the operator for building users, that is, the guiding price for building users to purchase electricity from the grid / operator, the first electricity sales price P. sellIt is the price signal that building users sell photovoltaic or energy storage back to operators or the grid. Building users decide whether to buy or sell the amount of electricity Q based on the first purchase price, the first sales price, load demand, and photovoltaic power generation. sale (t). The initial values ​​can be set based on historical electricity price levels and market strategies, for example, the purchase price of electricity is between 0.8 and 1.0 yuan / kWh, and the sales price of electricity is between 0.4 and 0.6 yuan / kWh. These two prices can be iteratively adjusted using optimization algorithms, such as genetic algorithms and particle swarm optimization, to maximize the energy sales revenue minus carbon emission costs under the electricity purchase distribution after building users respond to their pricing strategies, generating an upper-level pricing optimization model, i.e., maxZup=R. sale -Cco2, in each iteration, the upper-level pricing optimization model calculates the expected electricity purchase volume of building users based on the current electricity purchase and sale price, and updates the electricity sales revenue and carbon emission costs, while constraining the price within an allowable range, until the optimal balance between revenue and carbon emissions is achieved.

[0046] The lower-level energy optimization model takes building users as the decision-making body and uses minimizing building users' energy costs, maximizing overall satisfaction, and minimizing carbon emissions as the secondary optimization objectives. The energy cost (Cost) is calculated from the building users' electricity consumption, electricity purchase price, and electricity sales revenue at each time period. , where Q purchase (t) represents the actual electricity purchased at time t, Q sell (t) represents the electricity fed back to the grid. Overall satisfaction includes thermal comfort satisfaction and lighting environment satisfaction. Carbon emissions (Eco2) are calculated by multiplying the actual electricity purchased by building users by the corresponding carbon emission factor. An objective function is constructed by weighting and summing energy costs, overall satisfaction, and carbon emissions to generate a lower-level energy optimization model: minF = w1 × Cost - w2 × Total + w3 × Eco2. Overall satisfaction (Stotal) = w4 × Thermal comfort satisfaction + w5 × Lighting environment satisfaction. Here, w1, w2, and w3 are the weighting coefficients for each sub-objective of energy costs, overall satisfaction, and carbon emissions, while w4 and w5 are the weighting coefficients for thermal comfort satisfaction and lighting environment satisfaction. These can be set based on actual needs. w1 + w2 + w3 = 1, w4 + w5 = 1, and -w2 × Total indicates that the greater the overall satisfaction, the larger the negative value of this term, and the smaller the objective function F, thus ensuring that the direction of maximizing and minimizing satisfaction is consistent.

[0047] In the lower-level energy optimization model, indoor temperature comfort is constrained to ensure the thermal comfort of building users. Based on building design codes and user-defined comfort temperature ranges, the upper and lower limits Tmin and Tmax of the indoor temperature are extracted. For example, the office area is set at 22–26℃ as the adjustable temperature range. Then, based on the thermal inertia characteristics of the building envelope and the response speed of the air conditioning system, the maximum rate of change of indoor temperature per unit time, ΔTmax / Δt, is extracted. For example, air conditioning power limitations may limit the temperature adjustment to ±1.5℃ per hour. Based on the temperature range and rate of change, the building is treated as a virtual energy storage device: the adjustable indoor temperature range corresponds to the virtual capacity Cv, and the rate of change of indoor temperature within the allowable range corresponds to the virtual charging and discharging power Pv(t), where Cv = Tmax - Tmin, -P v max ≤dT(t) / dt≤P v max P v max =ΔTmax / Δt. During the optimization process, by controlling the air conditioning cooling / heating power, the indoor temperature change rate dT(t) / dt is ensured to not exceed the virtual charging / discharging power limit, while the temperature remains within the virtual capacity range, thus forming an indoor temperature comfort constraint. This indoor temperature comfort constraint can be directly integrated into the lower-level energy optimization model to ensure that the scheduling strategy satisfies energy efficiency optimization without exceeding the user's acceptable comfort range.

[0048] By nesting and iterating the upper-level pricing optimization model and the lower-level energy consumption optimization model, a two-layer optimization scheduling architecture is generated to achieve two-layer scheduling: the upper-level pricing optimization model outputs the electricity purchase and sale price signal, the lower-level energy consumption optimization model optimizes the energy consumption strategy based on the price signal, and feeds back the actual purchased electricity volume and load distribution. The upper-level pricing optimization model adjusts the pricing strategy based on the feedback, re-triggers the lower-level energy consumption optimization model to optimize, and iteratively generates the final scheduling strategy.

[0049] By constructing a two-layer optimized scheduling architecture, the energy operator's revenue maximization is coordinated with the building user's cost minimization, comfort optimization, and carbon emission control, providing a dynamic and implementable decision-making mechanism for the joint scheduling of multi-energy flows in smart buildings.

[0050] Furthermore, the thermal comfort satisfaction is quantified in the following way: the indoor temperature and humidity of the smart building are obtained to obtain a first environmental parameter; the first environmental parameter is substituted into the predicted average number of votes to obtain a first voting index; the first environmental parameter is substituted into the predicted percentage of dissatisfaction to obtain a first dissatisfaction index; and the thermal comfort satisfaction is calculated based on the first voting index and the first dissatisfaction index.

[0051] Specifically, indoor temperature and humidity are collected by sensors inside the building to form the first environmental parameter. This first environmental parameter is then substituted into the Predicted Average Voting Value (PMV) model. The PMV model is based on the human body heat balance equation and comprehensively considers six factors: air temperature, average radiant temperature, relative humidity, wind speed, human metabolic rate, and clothing thermal resistance. It outputs a continuous value ranging from -3 (cold) to +3 (hot), where 0 represents thermal neutrality, i.e., the most comfortable state. PMV(t) = f(Tin(t),RH(t),Va,Icl,M), where Tin(t) is the indoor temperature, RH(t) is the indoor humidity, Va is the air velocity, Icl is the clothing insulation value, and M is the metabolic rate. The first voting index is obtained through the PMV model. Simultaneously, the first environmental parameter is substituted into the Predicted Percentage of Dissatisfaction (PPD) indicator to quantify the proportion of people feeling uncomfortable in the current environment: PPD(t) = 100 - 95 × e -(0.03353⋅PMV(t)4+0.2179⋅PMV(t)2) The PPD (Percentage Point Dissatisfaction) value ranges from 5% to 100%, with lower PPD values ​​indicating higher thermal comfort. A first dissatisfaction index is calculated by predicting the percentage of dissatisfaction. Then, based on the first dissatisfaction index and the first dissatisfaction index, the thermal comfort satisfaction level S is calculated. thermal (t)=w PMV ×(3−|PMV(t)|) / 3+w PPD ×(100−PPD(t)) / 100, where w PMV w PPD Let w be the weighting coefficient. PMV +w PPD =1, used to balance the contribution of average human perception and the proportion of dissatisfaction. The calculation result is between 0 and 1, and the higher the thermal comfort satisfaction value, the more comfortable the indoor thermal environment.

[0052] For example, at 14:00 on April 16th, the indoor temperature was 25℃, humidity (RH) was 50%, air velocity (Va) was 0.1 m / s, clothing index (Icl) was 0.5clo, and metabolic rate (M) was 1.2met. Using the PMV formula, PMV = 0.3 (slightly warm). Substituting this into the PPD formula, PPD = 10.5%. PMV w PPD The values ​​are 0.6 and 0.4 respectively. The calculated thermal comfort satisfaction index is 0.6×(3-0.3) / 3+0.4×(100-10.5) / 100≈0.91, which reflects that the indoor environment is relatively comfortable.

[0053] By quantifying indoor temperature and humidity data into a thermal comfort satisfaction index, a numerical input that can be directly used as the objective function is provided for the lower-level energy optimization model. This enables building users' air conditioning control strategies to consider both energy saving and comfort, improving the accuracy of dual-level optimization scheduling and the overall thermal comfort satisfaction of the group.

[0054] Furthermore, the lighting environment satisfaction is quantified by: obtaining the total environmental illuminance of the smart building; substituting the total environmental illuminance into a Gaussian distribution function to calculate the lighting comfort index; and using the lighting comfort index as the lighting environment satisfaction.

[0055] Specifically, real-time total environmental illuminance data is obtained by illuminance sensors deployed in different functional areas of the building interior, such as office areas, corridors, and meeting rooms. The illuminance sensors are photodiode-type or digital illuminance meters, and are uploaded to the energy management system (EMS) through the building automation system (BMS) to form time-series data. The total environmental illuminance L(t) refers to the comprehensive light intensity after the superposition of multiple indoor lighting points and natural lighting, which is used to reflect the actual visual perception of the environment by the human body.

[0056] The lighting comfort index is calculated by substituting the total ambient illuminance L(t) into the Gaussian distribution function. The Gaussian function centers on the optimal comfortable illuminance range, Lopt, and the illuminance tolerance deviation σ characterizes the human body's sensitivity to light deviation. Its mathematical expression is: , where L opt Based on the building's usage scenario, such as 500 lx for office environments and 750 lx for meeting environments, σ ​​is generally set between 100 and 300 lx to reflect visual comfort tolerance. A larger value indicates a wider tolerance range, and vice versa. When the actual illuminance is close to the optimal illuminance, the output value approaches 1; the greater the deviation, the more the exponential decay approaches 0, thus achieving continuous quantification of lighting comfort. The lighting comfort index calculated using the Gaussian function is directly used as the lighting environment satisfaction index and input into the lower-level optimization model. This index, along with energy cost, thermal comfort, and carbon emissions, forms a multi-objective optimization function, thereby achieving coordinated optimization control between the building's visual environment and energy consumption.

[0057] For example, at 14:00 on April 16th, the total indoor illuminance of an office building was collected as L(t) = 620 lx. The optimal illuminance for the office area was set as Lopt = 500 lx, and the illuminance tolerance deviation σ = 150 lx. Then: f(t) = exp(-(620-500)). 2 / (2×150 2 ))≈0.726.

[0058] By mapping indoor illuminance signals to a standardized lighting comfort index, physical lighting parameters that are difficult to directly participate in optimization calculations are transformed into continuous evaluation indicators within the range of 0 to 1. This allows them to be incorporated into a multi-objective optimization function along with thermal comfort and economic indicators, thereby enabling the quantitative participation of the building lighting environment in the energy-saving control process to be optimized and improving both energy efficiency and indoor comfort.

[0059] The optimal scheduling scheme is obtained by performing optimization scheduling analysis on the photovoltaic power generation curve, the building electrical load curve, and the heating and cooling load curve through the two-layer optimization scheduling architecture.

[0060] During the target scheduling period, the smart building is optimized and scheduled according to the optimal scheduling scheme.

[0061] Furthermore, the optimal scheduling scheme is obtained by performing optimization scheduling analysis on the photovoltaic power generation curve, the building electricity load curve, and the heating and cooling load curve through the two-layer optimization scheduling architecture, including: updating the first electricity purchase price and the first electricity sales price through a genetic algorithm to obtain a first electricity price strategy; based on the first electricity price strategy, calling a mathematical programming solver to optimize the energy consumption strategy of the building users to obtain a first energy consumption strategy; iterating alternately between the first electricity price strategy and the first energy consumption strategy, and extracting the change in the objective function value between two consecutive iterations; when the change in the objective function value is less than a preset threshold, it is determined that a Nash equilibrium has been reached, and the optimal scheduling scheme is generated; wherein, the generation The optimal scheduling scheme is then derived, and further includes: extracting the equipment start-up and shutdown plan, energy storage charging and discharging plan, grid interaction power plan, flexible load response scheme, and indoor temperature setpoint from the optimal scheduling scheme; establishing a virtual charging and discharging power model and a virtual capacity model for the building's virtual energy storage device based on the thermal and cold inertia characteristics of the building envelope; substituting the indoor temperature setpoint into the virtual charging and discharging power model and the virtual capacity model to verify whether the indoor temperature setpoint meets the indoor temperature comfort constraint; if it does, outputting the equipment start-up and shutdown plan, the energy storage charging and discharging plan, the grid interaction power plan, and the flexible load response scheme as the optimal scheduling scheme.

[0062] Specifically, upper-level energy operators and lower-level building users approach Nash equilibrium through an alternating iterative process. First, based on the upper-level pricing optimization model in the two-level optimization scheduling architecture, the purchase price and sales price of electricity at each moment within the target scheduling period are arranged sequentially to form a one-dimensional electricity price vector. This vector is then encoded as chromosomes for a genetic algorithm, with each chromosome containing 2×T genes, where T is the number of hours in the scheduling period (e.g., 24 hours, resulting in a chromosome length of 48). Initially, the population size is set to 50-100, and the initial electricity price is randomly generated within an allowable range, such as a purchase price of 0.8-1.0 yuan / kWh and a sales price of 0.4-0.6 yuan / kWh. In each iteration, based on the current electricity price strategy, the purchase and sales volumes of building users after their response are predicted using the lower-level energy consumption optimization model, and then the upper-level objective function value is calculated. A tournament selection method is used, where k individuals (usually k=3) are randomly selected from the population each time, and the individual with the highest fitness is selected to enter the next generation. This process is repeated multiple times to generate a mating pool. A simulated binary crossover operator is used to perform a crossover operation on parent pairs in the mating pool with a crossover probability of 0.85. A polynomial mutation operator is used to perform a mutation operation on each newly generated offspring individual with a mutation probability of 1 / 48, generating a new generation of chromosomes. The new population generated after selection, crossover, and mutation is used as the next generation, and the above fitness evaluation and lower-level interaction process is repeated. When the improvement of the optimal fitness value in 15 consecutive generations is less than 10, the next generation is selected. -4 When the algorithm converges, it outputs the chromosome individual with the highest fitness in the current population as the first electricity price strategy. The first electricity price strategy includes the first electricity sales price sequence and the first electricity purchase price sequence for 24 hours.

[0063] After receiving the first electricity price strategy from the upper-level pricing optimization model, the lower-level energy optimization model uses the photovoltaic power generation, building electricity load, and heating and cooling load curves at each time step as known input data. Combining the first electricity price strategy, it calls the mathematical programming solver to optimize the energy consumption strategy of building users and generates the first energy consumption strategy. The first energy consumption strategy includes the user's electricity purchase power sequence, energy storage charging and discharging power sequence, air conditioning power sequence, etc., and performs joint optimization with energy cost, comfort, and carbon emissions as constraints or objectives.

[0064] In the two-layer interaction process, strategy coupling optimization is achieved through an alternating iterative mechanism: after obtaining the first energy consumption strategy, it is returned to the upper-layer pricing optimization model. The upper-layer pricing optimization model adjusts the electricity price strategy based on the current feedback of the first energy consumption strategy, and the lower-layer energy consumption optimization model re-optimizes the energy consumption strategy based on the updated electricity price. After each iteration, the fitness value of the upper-layer operator, i.e., the upper-layer objective function value, is calculated. By extracting the change in the objective function value between two consecutive iterations, when this change is less than a preset convergence threshold, such as 0.001, it is determined that the current combination of electricity price strategy and energy consumption strategy has reached a Nash equilibrium state, i.e., the upper and lower layer strategies no longer change significantly. The iteration stops, and the current optimal electricity price strategy and the corresponding optimal energy consumption strategy combination are output as the optimal scheduling scheme.

[0065] After generating the optimal scheduling scheme, a structured analysis is performed on the optimal scheduling scheme, and the key control variables are decomposed according to the time series. Specifically, this includes extracting equipment start-up and shutdown plans, energy storage charging and discharging plans, grid interaction power plans, flexible load response schemes, and indoor temperature setpoints. Among them, the equipment start-up and shutdown state vectors are such as the 0 / 1 operation flags of air conditioning, lighting, and controllable loads; the energy storage charging and discharging plans are such as the charging power, discharging power, and state of charge curves of batteries at each time; the grid interaction power plans are such as the hourly sequence of the power purchased by users from operators and the power purchased from the grid; the flexible load response scheme is the scheduling period and power adjustment amount of interruptible loads or loads that can be shifted; and the indoor temperature setpoint is the target temperature at each time derived from the air conditioning power in the lower-level optimization.

[0066] Based on the thermal and cold inertia characteristics of the building envelope, a virtual charging / discharging power model and a virtual capacity model for the building's virtual energy storage device are established. The adjustable range of indoor temperature deviating from the comfort zone is defined as the virtual capacity, used to characterize the acceptable temperature fluctuation space of the building. Simultaneously, based on the allowable fluctuation range and maximum allowable rate of change of indoor temperature, the adjustable rate of indoor temperature at each moment is extracted and mapped to the virtual charging / discharging power model Pv(t). The sequence of indoor temperature setpoints from candidate scheduling schemes is input into the virtual energy storage model, and the temperature change rate is calculated at each moment to determine whether |dT(t) / dt|≤P. v maxThe system verifies whether the temperature trajectory always satisfies the comfort constraint range of Tmin ≤ T(t) ≤ Tmax. If the constraint is violated at any moment, the scheduling scheme is deemed unexecutable in the thermal comfort dimension, triggering a revision or rollback of the optimization results. If the constraint is satisfied throughout the entire time period, it indicates that the temperature setting strategy is achievable within the building's thermal dynamic response capability. After completing the comfort and physical feasibility verification, the verified equipment start-up and shutdown plans, energy storage charging and discharging plans, grid interaction power plans, and flexible load response schemes are compiled into a final execution instruction set and output as the optimal scheduling scheme. This set is then executed in real time by the building energy management system within the target scheduling cycle, thereby achieving multi-energy flow collaborative optimization control and stable operation under comfort constraints.

[0067] By constructing a two-layer game optimization mechanism involving price guidance, load response, and iterative convergence, the decision-making processes of energy operators and building users are collaboratively coupled, enabling global optimal matching of photovoltaic output, building load, and heating and cooling demand within a unified framework. The stability of the strategy is ensured through Nash equilibrium, while indoor temperature constraints are verified by combining a virtual energy storage model. This ensures that the final scheduling scheme achieves optimal coordination between economy, low carbon emissions, and comfort, thereby improving the energy utilization efficiency and indoor comfort of smart buildings.

[0068] Example 2, based on the same inventive concept as the AI-based smart building multi-energy flow joint optimization scheduling method in the previous examples, such as... Figure 3 As shown, this application provides an AI-based smart building multi-energy flow joint optimization scheduling system, wherein the AI-based smart building multi-energy flow joint optimization scheduling system includes: The dataset construction module 11 is used to clean and normalize the historical multi-source data of the smart building to construct a differentiated training dataset; the prediction curve acquisition module 12 is used to analyze the differentiated training dataset to obtain the target hourly prediction curve of the smart building in the target scheduling period, wherein the target hourly prediction curve includes the photovoltaic power generation curve, the building electrical load curve, and the heating and cooling load curve; the scheduling architecture construction module 13 is used to construct a two-layer optimized scheduling architecture with the energy operator as the leader and the building user as the follower; the optimized scheduling analysis module 14 is used to perform optimized scheduling analysis on the photovoltaic power generation curve, the building electrical load curve, and the heating and cooling load curve through the two-layer optimized scheduling architecture to obtain the optimal scheduling scheme; the optimized scheduling execution module 15 is used to perform optimized scheduling execution on the smart building according to the optimal scheduling scheme in the target scheduling period.

[0069] Furthermore, the dataset construction module 11 is also used to: extract a first historical data group from the historical multi-source data, and perform outlier removal and dimension unification on the first historical data group to obtain a first standardized dataset; divide the first standardized dataset based on the four-season division strategy to obtain historical division results; and combine the spring dataset, summer dataset, autumn dataset, and winter dataset in the historical division results to form the differentiated training dataset; wherein, the first historical data group includes first historical meteorological data, first historical building internal environment data, first historical multi-energy load data, and first historical distributed power output data.

[0070] Furthermore, the prediction curve acquisition module 12 is also used to: initialize the first initial weights and the first initial threshold of the backpropagation neural network; use the first initial weights and the first initial threshold as optimization objects, and obtain the first target weights and the first target threshold through particle swarm iteration calculation; substitute the first target weights and the first target threshold into the backpropagation neural network to construct the first prediction network, the second prediction network and the third prediction network respectively; input the differentiated training dataset into the first prediction network, the second prediction network and the third prediction network respectively, and obtain the photovoltaic power generation curve, the building electrical load curve and the heating and cooling load curve through forward propagation calculation.

[0071] Furthermore, the prediction curve acquisition module 12 is also used to: extract the solar radiation intensity, ambient temperature and photovoltaic power generation at the same time the day before the prediction date as the first input parameters; obtain the solar radiation intensity and ambient temperature at the same time the prediction date as the second input parameters; and concatenate the first input parameters and the second input parameters into the first prediction network to obtain the photovoltaic power generation curve through forward propagation calculation.

[0072] Furthermore, the scheduling architecture construction module 13 is also used to: take maximizing the difference between the energy operator's energy sales revenue and carbon emission costs as the first optimization objective, and take the first electricity purchase price and the first electricity sales price set for the building users as the first decision variables to generate an upper-level pricing optimization model; take minimizing the building users' energy costs, maximizing overall satisfaction, and minimizing carbon emissions as the second optimization objective to generate a lower-level energy consumption optimization model; wherein, the overall satisfaction includes thermal comfort satisfaction and lighting environment satisfaction; and construct the two-layer optimized scheduling architecture based on the upper-level pricing optimization model and the lower-level energy consumption optimization model.

[0073] Furthermore, the scheduling architecture construction module 13 is also used to: obtain the indoor temperature and indoor humidity of the smart building to obtain a first environmental parameter; substitute the first environmental parameter into the predicted average vote count index to obtain a first voting index; substitute the first environmental parameter into the predicted dissatisfaction percentage index to obtain a first dissatisfaction index; and calculate the thermal comfort satisfaction based on the first voting index and the first dissatisfaction index.

[0074] Furthermore, the scheduling architecture construction module 13 is also used to: obtain the total environmental illuminance of the smart building; substitute the total environmental illuminance into a Gaussian distribution function to calculate the lighting comfort index; and use the lighting comfort index as the lighting environment satisfaction.

[0075] Furthermore, the scheduling architecture construction module 13 is also used for: the constraints of the lower-level energy consumption optimization model include indoor temperature comfort constraints, including: extracting the rate of change of indoor temperature within the allowable range and constructing virtual charging and discharging power; extracting the upper and lower limits of adjustable indoor temperature and constructing virtual capacity; and generating the indoor temperature comfort constraints based on the virtual charging and discharging power and the virtual capacity.

[0076] Furthermore, the optimized scheduling analysis module 14 is also used to: update the first electricity purchase price and the first electricity sales price through a genetic algorithm to obtain a first electricity price strategy; based on the first electricity price strategy, call a mathematical programming solver to optimize the energy consumption strategy of the building users to obtain a first energy consumption strategy; perform alternating iterations on the first electricity price strategy and the first energy consumption strategy, and extract the change in the objective function value between two consecutive iterations; when the change in the objective function value is less than a preset threshold, determine that a Nash equilibrium has been reached, and generate the optimal scheduling scheme; wherein, after generating the optimal scheduling scheme, the module further includes: extracting the optimal scheduling scheme. The optimal scheduling scheme includes equipment start-up and shutdown plans, energy storage charging and discharging plans, grid interaction power plans, flexible load response schemes, and indoor temperature setpoints. Based on the thermal and cold inertia characteristics of the building envelope, a virtual charging and discharging power model and a virtual capacity model for the building's virtual energy storage equipment are established. The indoor temperature setpoints are substituted into the virtual charging and discharging power models and the virtual capacity models to verify whether the indoor temperature setpoints meet the indoor temperature comfort constraints. If they do, the equipment start-up and shutdown plans, the energy storage charging and discharging plans, the grid interaction power plans, and the flexible load response schemes are output as the optimal scheduling scheme.

[0077] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The AI-based smart building multi-energy flow joint optimization scheduling method and specific examples in the aforementioned embodiment one are also applicable to the AI-based smart building multi-energy flow joint optimization scheduling system in this embodiment. Through the foregoing detailed description of the AI-based smart building multi-energy flow joint optimization scheduling method, those skilled in the art can clearly understand the AI-based smart building multi-energy flow joint optimization scheduling system in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0078] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0079] Obviously, those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A multi-energy flow joint optimization scheduling method for smart buildings based on artificial intelligence, characterized in that, include: Historical multi-source data of smart buildings are cleaned and normalized to construct a differentiated training dataset; The differential training dataset is analyzed to obtain the target hourly prediction curve of the smart building in the target scheduling cycle, wherein the target hourly prediction curve includes the photovoltaic power generation curve, the building electrical load curve, and the cooling and heating load curve; Construct a two-tier optimized scheduling architecture with energy operators as leaders and building users as followers; The optimal scheduling scheme is obtained by performing optimization scheduling analysis on the photovoltaic power generation curve, the building electrical load curve, and the heating and cooling load curve through the two-layer optimization scheduling architecture. During the target scheduling period, the smart building is optimized and scheduled according to the optimal scheduling scheme.

2. The AI-based smart building multi-energy flow joint optimization scheduling method as described in claim 1, characterized in that, Historical multi-source data on smart buildings are cleaned and normalized to construct a differentiated training dataset, including: Extract the first historical data group from the historical multi-source data, and perform outlier removal and unit unification on the first historical data group to obtain the first standardized dataset. The first standardized dataset was divided based on the four-season division strategy to obtain historical division results; The differential training dataset is formed by combining the spring, summer, autumn, and winter datasets from the historical segmentation results. The first historical data group includes first historical meteorological data, first historical building internal environment data, first historical multi-energy load data, and first historical distributed power output data.

3. The AI-based smart building multi-energy flow joint optimization scheduling method as described in claim 2, characterized in that, Analyzing the differential training dataset yields the hourly prediction curves of the smart building during the target scheduling cycle, including: Initialize the first initial weights and the first initial threshold of the backpropagation neural network; Using the first initial weight and the first initial threshold as the optimization objects, the first target weight and the first target threshold are obtained by particle swarm iteration calculation; Substitute the first target weight and the first target threshold into the backpropagation neural network to construct the first prediction network, the second prediction network and the third prediction network respectively. The differentiated training dataset is input into the first prediction network, the second prediction network, and the third prediction network, respectively, and the photovoltaic power generation curve, the building electrical load curve, and the heating and cooling load curve are obtained through forward propagation.

4. The AI-based smart building multi-energy flow joint optimization scheduling method as described in claim 3, characterized in that, The differential training dataset is input into the first prediction network, and the photovoltaic power generation curve is obtained through forward propagation, including: Extract the solar radiation intensity, ambient temperature, and photovoltaic power generation at the same time the day before the forecast date as the first input parameters; The solar radiation intensity and ambient temperature at the same time on the predicted date are obtained as the second input parameters; The first input parameter and the second input parameter are concatenated and input into the first prediction network, and the photovoltaic power generation curve is obtained by forward propagation calculation.

5. The AI-based smart building multi-energy flow joint optimization scheduling method as described in claim 1, characterized in that, Construct a two-tier optimized scheduling architecture with energy operators as leaders and building users as followers, including: With maximizing the difference between the energy operator's energy sales revenue and carbon emission costs as the first optimization objective, and using the first electricity purchase price and the first electricity sales price set for the building users as the first decision variables, an upper-level pricing optimization model is generated. The lower-level energy optimization model is generated with the second optimization objective being to minimize the energy cost, maximize the overall satisfaction, and minimize the carbon emissions of the building users. The overall satisfaction includes thermal comfort satisfaction and lighting environment satisfaction. The two-layer optimized scheduling architecture is constructed based on the upper-layer pricing optimization model and the lower-layer energy consumption optimization model.

6. The AI-based smart building multi-energy flow joint optimization scheduling method as described in claim 5, characterized in that, The thermal comfort satisfaction level is quantified in the following way: The indoor temperature and humidity of the smart building are obtained to obtain the first environmental parameter; Substituting the first environmental parameter into the predicted average vote count index, we obtain the first voting index; Substituting the first environmental parameter into the predicted percentage of dissatisfaction, we obtain the first dissatisfaction index; The thermal comfort satisfaction level is calculated based on the first voting index and the first dissatisfaction index.

7. The AI-based smart building multi-energy flow joint optimization scheduling method as described in claim 5, characterized in that, The satisfaction level with the lighting environment was quantified in the following way: Obtain the total ambient illuminance of the smart building; Substituting the total environmental illuminance into the Gaussian distribution function, the lighting comfort index is calculated; The lighting comfort index is used as the satisfaction level of the lighting environment.

8. The AI-based smart building multi-energy flow joint optimization scheduling method as described in claim 5, characterized in that, The constraints of the lower-level energy optimization model include indoor temperature comfort constraints, including: Extract the rate of change of indoor temperature within the allowable range and construct a virtual charging and discharging power; Extract the upper and lower limits of the adjustable indoor temperature to construct a virtual capacity; The indoor temperature comfort constraint is generated based on the virtual charging and discharging power and the virtual capacity.

9. The AI-based smart building multi-energy flow joint optimization scheduling method as described in claim 8, characterized in that, The optimal scheduling scheme is obtained by performing optimization scheduling analysis on the photovoltaic power generation curve, the building electrical load curve, and the heating and cooling load curve using the two-layer optimized scheduling architecture, including: The first electricity purchase price and the first electricity sales price are updated using a genetic algorithm to obtain the first electricity pricing strategy; Based on the first electricity price strategy, the mathematical programming solver is invoked to optimize the energy consumption strategy of the building users, and the first energy consumption strategy is obtained. The first electricity pricing strategy and the first energy consumption strategy are iterated alternately to extract the change in the objective function value between two consecutive iterations. When the change in the objective function value is less than a preset threshold, it is determined that Nash equilibrium has been reached, and the optimal scheduling scheme is generated. The process of generating the optimal scheduling scheme further includes: Extract the equipment start-up and shutdown plan, energy storage charging and discharging plan, grid interaction power plan, flexible load response plan, and indoor temperature setpoint from the optimal scheduling scheme; Based on the thermal and cold inertia characteristics of building envelope, a virtual charging and discharging power model and a virtual capacity model of building virtual energy storage device are established. Substitute the indoor temperature setpoint into the virtual charge / discharge power model and the virtual capacity model to verify whether the indoor temperature setpoint meets the indoor temperature comfort constraint. If satisfied, the equipment start-up and shutdown plan, the energy storage charging and discharging plan, the grid interaction power plan, and the flexible load response scheme will be output as the optimal scheduling scheme.

10. A smart building multi-energy flow joint optimization scheduling system based on artificial intelligence, characterized in that, The steps for implementing the AI-based smart building multi-energy flow joint optimization scheduling method according to any one of claims 1 to 9, wherein the AI-based smart building multi-energy flow joint optimization scheduling system comprises: The dataset construction module is used to clean and normalize the historical multi-source data of smart buildings to build a differentiated training dataset. The prediction curve acquisition module is used to analyze the differentiated training dataset to obtain the target hourly prediction curve of the smart building in the target scheduling cycle, wherein the target hourly prediction curve includes the photovoltaic power generation curve, the building electrical load curve, and the cooling and heating load curve; The scheduling architecture building module is used to construct a two-tier optimized scheduling architecture with energy operators as leaders and building users as followers. The optimized scheduling analysis module is used to perform optimized scheduling analysis on the photovoltaic power generation curve, the building electrical load curve, and the heating and cooling load curve through the two-layer optimized scheduling architecture to obtain the optimal scheduling scheme; An optimized scheduling execution module is used to optimize the scheduling execution of the smart building according to the optimal scheduling scheme during the target scheduling period.

Images