Building group flexible load coordinated regulation and control method and device based on artificial intelligence

Abstract

The invention provides a building group flexible load cooperative regulation and control method and device based on artificial intelligence, and relates to the technical field of artificial intelligence, and the method comprises the steps: obtaining multi-dimensional feature data of a target building; taking the multi-dimensional feature data as input, and outputting future multi-time scale load prediction data of at least one load unit of the target building through a pre-trained load prediction model; based on the load prediction data of the multiple time scales, a pre-constructed target function is optimized and solved through a target optimization algorithm improved based on the coupling relation of all power consumption devices in the target building, a cooperative regulation and control strategy of at least one load unit is obtained, and the cooperative regulation and control strategy comprises control parameters of the corresponding load unit; and executing the coordinated regulation and control strategy on the at least one load unit, and dynamically adjusting the coordinated regulation and control strategy according to the real-time operation state of the at least one load unit. According to the invention, the regulation and control efficiency and precision of the building flexible load are improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to an artificial intelligence-based method for coordinated control of flexible loads in building complexes, an artificial intelligence-based device for coordinated control of flexible loads in building complexes, a machine-readable storage medium, and a terminal device. Background Technology

[0002] With the expansion of building complexes and the rise in energy demand, flexible loads such as air conditioning, lighting, and elevators now account for over 60% of total building energy consumption, becoming the core focus of building energy conservation and load management. Currently, existing building flexible load control methods mostly employ setpoint control or single-parameter feedback control modes, which have the following key drawbacks: 1. Insufficient prediction accuracy: Traditional control methods rely on historical statistical data or simple linear prediction models, which do not fully integrate dynamic characteristics such as weather, population flow, and building energy consumption habits, resulting in a prediction error of more than 20% in load change trends. They are unable to adapt to complex scenarios such as seasonal changes and the switching between weekdays and holidays.

[0003] 2. Lack of coordination: Existing technologies usually treat loads such as air conditioning, lighting, and elevators as independent objects and regulate them separately, ignoring the coupling relationship between loads (such as the effect of air conditioning load changes on indoor temperature, which in turn affects the lighting brightness demand), which can easily lead to regulation conflicts, such as the problem of excessive cooling by air conditioning and excessive energy consumption by lighting.

[0004] 3. Delayed response: The passive logic of "monitoring-feedback-control" is adopted, and the response to external incentives such as fluctuations in grid peak and valley electricity prices and changes in renewable energy output is delayed by more than 5 minutes, making it impossible to achieve precise matching between peak shaving and valley filling and priority consumption of green electricity.

[0005] 4. Imbalance between comfort and energy efficiency: In pursuit of energy-saving effects, excessive restrictions on load operation, such as setting the air conditioner temperature too high in summer, which leads to a decrease in people's comfort, or setting the elevator dispatch interval too long, which affects traffic efficiency.

[0006] Therefore, in order to address the above problems, there is an urgent need for a flexible load control technology that integrates high-precision forecasting, multi-load coordination, rapid response, and takes into account both energy efficiency and comfort, so as to improve the energy management level of building complexes. Summary of the Invention

[0007] The purpose of this application is to provide an artificial intelligence-based method for coordinated control of flexible loads in building complexes, an artificial intelligence-based device for coordinated control of flexible loads in building complexes, a machine-readable storage medium, and a terminal device to solve the above-mentioned problems.

[0008] To achieve the above objectives, the first aspect of this application provides a method for coordinated control of flexible loads in building complexes based on artificial intelligence, comprising: Acquire multidimensional feature data of the target building, including static attribute parameters, dynamic operating parameters, external environment parameters, and user behavior parameters of the target building; Using the multidimensional feature data as input, the pre-trained load prediction model outputs load prediction data for at least one load unit of the target building in the future at multiple time scales. Based on the load forecast data at multiple time scales, the pre-constructed objective function is optimized and solved by an improved objective optimization algorithm based on the coupling relationship of each power consumption device in the target building, thereby obtaining the coordinated control strategy of at least one load unit, wherein the coordinated control strategy includes the control parameters of the corresponding load unit. The coordinated control strategy is executed on the at least one load unit, and the coordinated control strategy is dynamically adjusted according to the real-time operating status of the at least one load unit.

[0009] Optionally, the static attribute parameters include the building structure parameters of the target building, the rated parameters of each power consumption device, and building information model data, wherein the building information model data includes spatial topology data representing the spatial relationships of different load units in the target building; The dynamic operating parameters include at least one of the following: real-time operating data of each power consumption device in the area corresponding to each load unit, temperature and humidity, CO2 concentration, light intensity, and output data of the distributed photovoltaic / energy storage equipment of the target building. The external environmental parameters include at least one of the following: meteorological data, electricity price data, and photovoltaic power output forecast data for future periods; The user behavior parameters include at least one of the following: the pedestrian density in the area corresponding to each load unit in the target building, the number of users making reservations in the area corresponding to each load unit, and user feedback on the comfort level of the area corresponding to each load unit.

[0010] Optionally, the load forecasting model includes: Serial spatiotemporal graph neural network model and time sequence neural network model; The spatiotemporal graph neural network model is constructed based on the spatial topology data of the target building, and the temporal neural network model is a long short-term memory neural network model. The spatiotemporal graph neural network model is used to extract the spatial correlation features of each load unit based on the multidimensional feature data, and the extracted spatial correlation features are input into the long short-term memory neural network model. The long short-term memory neural network model is used to extract time-series features representing the time-varying patterns of each load unit based on the received spatial correlation features, and to generate load prediction data for each load unit at multiple time scales in the future based on the extracted time-series features.

[0011] Optionally, the training process of the load prediction model includes: Obtain historical multidimensional feature data of other buildings, and train the load prediction model at each time scale using the historical multidimensional feature data of other buildings to determine the model parameters of the load prediction model; The historical multidimensional feature data of the target building is obtained, and the load prediction model is trained at each time scale using the historical multidimensional feature data of the target building to update the model parameters of the load prediction model. When training the load prediction model at different time scales, the weights of static attribute parameters, dynamic operating parameters, external environment parameters, or user behavior parameters in the input historical multidimensional feature data are dynamically adjusted based on the current training time scale.

[0012] Optionally, the multiple time scales include at least: First time scale, second time scale, and third time scale; The duration of the first time scale is shorter than that of the second time scale, and the duration of the second time scale is shorter than that of the third time scale; The weights of static attribute parameters, dynamic runtime parameters, external environment parameters, or user behavior parameters in the input historical multidimensional feature data are dynamically adjusted based on the current training timescale, including: If the current training timescale is the first timescale, increase the weights of dynamic running parameters and user behavior parameters in the input historical multidimensional feature data; If the current training timescale is the second timescale, increase the weights of the dynamic running parameters and external environment parameters in the input historical multidimensional feature data, and the weights of the dynamic running parameters and external environment parameters are the same. If the current training timescale is the third timescale, increase the weight of the external environment parameters in the input historical multidimensional feature data.

[0013] Optionally, output load forecast data for at least one load unit of the target building over multiple time scales in the future, including: If the deviation between the load forecast data of at least one load unit of the target building in the first time scale and its load forecast data in the corresponding time period of the second time scale is greater than a preset first deviation threshold, the load forecast data of at least one load unit of the target building in the first time scale is updated with the load forecast data in the corresponding time period of the second time scale. If the deviation between the load forecast data of at least one load unit of the target building in the second time scale and its load forecast data in the corresponding time period of the third time scale is greater than a preset second deviation threshold, the load forecast data of at least one load unit of the target building in the second time scale shall be updated with the load forecast data in the corresponding time period of the third time scale.

[0014] Optionally, output load forecast data for at least one load unit of the target building over multiple time scales in the future, including: For each time scale: Obtain the load forecast data and the actual load measurement data of the previous time step; The correction factor for the current load forecast data is determined based on the error between the load forecast data and the measured load data of the previous time. The load forecast data at the current moment is corrected using the aforementioned correction coefficient.

[0015] Optionally, the correction factor for the current load forecast data is determined based on the error between the load forecast data and the measured load data of the previous time step, including: The correction factor is determined using the following formula:

[0016] in, Let be the correction factor at time t. The correction factor at time t-1 For smoothing coefficients, The load measurement data is at time t-1. This is the load forecast data at time t-1.

[0017] Optionally, the objective function is:

[0018] in, Let be the energy consumption function. For cost function, For comfort function, Let be the power grid response function. The weights of the energy consumption function, The weights of the cost function, The weights of the comfort function, The weights are those of the grid response function; the energy consumption function, cost function, comfort function, and grid response function respectively represent the mapping relationship between the energy consumption, cost, comfort value, and grid response value of the target building and the control parameters of each load unit. The pre-constructed objective function is optimized and solved using an improved objective optimization algorithm based on the coupling relationship of various power consumption devices in the target building, including: S1. Construct a coupling factor to represent the coupling relationship between each power consumption device. The coupling factor is used to represent the degree of influence of the operation of any power consumption device on the control parameters of other power consumption devices. S2. Initialize the population by encoding the control parameters and coupling factors of each load unit into an individual, with each individual representing a solution to the objective function. S3. For each individual in the population, calculate the fitness value of the current individual using the objective function; S4. Determine the non-dominant relationship of each individual based on its fitness value, and then rank each individual in a hierarchical manner according to its non-dominant relationship. S5. Based on the hierarchical sorting results of each individual, determine the parent individuals, perform crossover and mutation operations on the parent individuals to update the individuals in the current population. If the convergence condition is met, take the individual with the best fitness value in the current population as the optimal solution of the objective function, and use the optimal solution as the collaborative control strategy of the at least one load unit; otherwise, return to step S3.

[0019] Optionally, the coordinated control strategy is dynamically adjusted based on the real-time operating status of the at least one load unit, including: Real-time acquisition of dynamic operating parameters, external environmental parameters, and user behavior parameters for each load unit corresponding to the target building; If, through a pre-built conflict resolution knowledge base, it is determined that at least one of the dynamic operating parameters, external environment parameters, and user behavior parameters of the at least one load unit meets the conflict conditions, the corresponding conflict resolution rules are determined through the conflict resolution knowledge base, and the current collaborative control strategy is adjusted according to the conflict resolution rules. The conflict resolution knowledge base includes at least conflict resolution rules corresponding to different conflict conditions. The conflict conditions include at least one of the following: dynamic operating parameters of at least one load unit, external environment parameters, and user behavior parameters, which must meet preset conditions.

[0020] A second aspect of this application provides an artificial intelligence-based flexible load coordination and control device for building complexes, comprising: The data acquisition module is configured to acquire multidimensional feature data of the target building, including static attribute parameters, dynamic operating parameters, external environment parameters, and user behavior parameters of the target building. The load forecasting module is configured to take the multidimensional feature data as input and output load forecasting data for at least one load unit of the target building in the future at multiple time scales through a pre-trained load forecasting model. The coordinated control strategy module is configured to optimize and solve a pre-constructed objective function based on the load forecast data of the multi-time scale using an improved objective optimization algorithm based on the coupling relationship of each power consumption device in the target building, so as to obtain the coordinated control strategy of the at least one load unit. The coordinated control strategy includes the control parameters of the corresponding load unit. The coordinated control module is configured to execute the coordinated control strategy on the at least one load unit and dynamically adjust the coordinated control strategy according to the real-time operating status of the at least one load unit.

[0021] In a third aspect, this application provides a machine-readable storage medium storing instructions that, when executed by a processor, configure the processor to perform the artificial intelligence-based flexible load coordination control method for building clusters as described above.

[0022] In a fourth aspect, this application provides a terminal device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the artificial intelligence-based flexible load coordinated control method for building clusters as described above.

[0023] This application predicts the load forecast data of each load unit of the target building by integrating the static attribute parameters, dynamic operating parameters, external environmental parameters and user behavior parameters of the target building. Based on the coupling relationship of each load unit, an improved target optimization algorithm is constructed. Based on the obtained load forecast data and the target optimization algorithm, the pre-constructed objective function is optimized and solved to obtain the coordinated control strategy of each load unit. This can overcome the defects of low precision, poor coordination, lag response and imbalance between energy efficiency and comfort in the existing technology, and improve the control efficiency and precision of building flexible load.

[0024] Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0025] The accompanying drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the following detailed description to explain the embodiments of this application, but do not constitute a limitation on the embodiments of this application. In the drawings: Figure 1 A flowchart illustrating the method for coordinated control of flexible loads in building complexes based on artificial intelligence, provided in a preferred embodiment of this application. Figure 2 A model structure diagram of the load forecasting model provided in the preferred embodiment of this application; Figure 3A schematic diagram of an artificial intelligence-based flexible load collaborative control device for building complexes, provided as a preferred embodiment of this application; Figure 4 A schematic diagram of a terminal device provided for a preferred embodiment of this application.

[0026] Explanation of reference numerals in the attached figures 10 - Terminal device, 100 - Processor, 101 - Memory, 102 - Computer program. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0028] It should be noted that the technical solutions of the various embodiments of this application can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by this application.

[0029] To solve the above problems, such as Figure 1 As shown, the first aspect of this application provides a method for coordinated control of flexible loads in building complexes based on artificial intelligence, comprising: S100. Obtain multi-dimensional feature data of the target building. The multi-dimensional feature data includes the static attribute parameters, dynamic operation parameters, external environment parameters, and user behavior parameters of the target building. S200: Using multidimensional feature data as input, the pre-trained load forecasting model outputs load forecasting data for at least one load unit of the target building in the future at multiple time scales. S300: Based on load forecast data at multiple time scales, the pre-constructed objective function is optimized and solved by an improved objective optimization algorithm based on the coupling relationship of each power consumption device in the target building, so as to obtain a coordinated control strategy for at least one load unit. The coordinated control strategy includes the control parameters of the corresponding load unit. S400: Execute a coordinated control strategy for at least one load unit, and dynamically adjust the coordinated control strategy according to the real-time operating status of at least one load unit.

[0030] Thus, this application predicts the load forecast data of each load unit of the target building by integrating the static attribute parameters, dynamic operating parameters, external environmental parameters and user behavior parameters of the target building, and constructs an improved target optimization algorithm based on the coupling relationship of each load unit. Based on the obtained load forecast data and the target optimization algorithm, the pre-constructed objective function is optimized and solved to obtain the coordinated control strategy of each load unit. This can overcome the defects of low precision, poor coordination, lag response and imbalance between energy efficiency and comfort in the existing technology, and improve the control efficiency and precision of building flexible load.

[0031] In step S100, this application first constructs a four-dimensional feature system that integrates "static parameters, dynamic operation, external environment, and user behavior", and then achieves full data collection through a distributed sensing network.

[0032] The static attribute parameters include the building structure parameters of the target building, the rated parameters of each power-consuming device, and building information model (BIM) data. The BIM data includes spatial topology data representing the spatial relationships between different load units within the target building. For example, building structure parameters include, but are not limited to, building floor height, wall insulation coefficient, and window area; rated parameters of power-consuming devices include, but are not limited to, rated cooling / heating capacity of air conditioners, power of lighting equipment such as individual lamps, and rated load and speed of elevators; the BIM data can be obtained from spatial layout data exported from the target building's BIM model. This exported data is stored offline in a local database and updated at a preset period, such as quarterly.

[0033] Dynamic operating parameters include at least one of the following: real-time operating data of each power-consuming device in the corresponding area of ​​each load unit, temperature and humidity, CO2 concentration, light intensity, and output data of the distributed photovoltaic / energy storage equipment in the target building. It is understood that load units can be divided according to the function of the target building or according to floors. For example, the target building can be divided into different areas according to function, and a load unit can be multiple areas with similar functions on one floor of the target building, or each floor can be divided into one load unit; this is not limited here. The real-time operating data of each power-consuming device includes, but is not limited to, the temperature of the air conditioner evaporator (accuracy ±0.1℃), fan speed, real-time power of each lighting device (accuracy ±0.5%), elevator operating current and stopping floors, etc.; simultaneously, indoor temperature and humidity (±0.3℃ / ±3%RH), CO2 concentration (±50ppm), light intensity (±1lux), etc., of each load unit can be collected through environmental sensors; the output data of the distributed photovoltaic / energy storage equipment in the target building can be obtained through the grid interface, collecting real-time voltage, current, and distributed photovoltaic / energy storage output data, with a sampling frequency set to 10s / time.

[0034] External environmental parameters include at least one of the following: meteorological data, electricity price data, and photovoltaic power output forecast data for future periods. For example, hourly meteorological data (temperature, humidity, sunshine, precipitation) for the next 7 days can be obtained by connecting to a meteorological platform via API, peak-valley-flat electricity price periods and price standards can be obtained by connecting to a power grid marketing system, and photovoltaic power output forecast data can be obtained by connecting to a new energy monitoring platform. The update frequency can be set to 15 minutes / time.

[0035] User behavior parameters include at least one of the following: pedestrian density in the corresponding area of ​​each load unit in the target building, the number of users making reservations in the corresponding area of ​​each load unit, and user feedback on comfort in the corresponding area of ​​each load unit. For example, pedestrian density on each floor can be collected by infrared array sensors (detection distance 0-5m, accuracy ±0.1 people / ㎡), and data such as meeting reservations (time, number of people, location), merchant business hours, and holiday schedules can be obtained through office systems / shopping mall management platforms. User comfort feedback (temperature, brightness, elevator waiting satisfaction, rating range 1-5 points) can be collected through mobile devices, and the update frequency can be set to once every 5 minutes.

[0036] In this application, after collecting multi-dimensional data of the target building, further noise reduction, normalization, and feature enhancement processing are performed on the collected multi-dimensional data, specifically including: (1) Abnormal data cleaning: The isolated forest algorithm is used to identify abrupt changes in data (such as temperature jumps caused by sensor failures), and the 3σ criterion is used to remove outliers that deviate from the normal range. The accuracy of abnormal data processing is ≥98%. For missing data, interpolation based on spatiotemporal correlation is used to complete the data, and the completion error is ≤3%.

[0037] (2) Data normalization: The Z-Score normalization method is adopted to normalize the features of different dimensions (such as temperature °C, power kW, and population density people / ㎡) to the [0,1] interval, thereby eliminating the influence of the difference in dimensions on model training.

[0038] (3) Spatiotemporal feature fusion: Construct an attention mechanism fusion model, highlighting the feature weights of key time nodes such as peak and valley periods and commuting peaks in the time dimension; strengthening the feature weights of densely populated areas (such as the first floor of a commercial complex and conference rooms in office areas) and densely populated areas in the spatial dimension, generating a feature vector set containing spatiotemporal correlation information with a dimension of 64 dimensions / sample.

[0039] The data processing described above can be implemented using existing algorithms, and no restrictions are imposed here.

[0040] In step S200, the load forecasting model of this application includes: a cascaded spatiotemporal graph neural network model and a time-series neural network model. The spatiotemporal graph neural network model is constructed based on the spatial topology data of the target building, and the time-series neural network model is a long short-term memory neural network model. The spatiotemporal graph neural network model is used to extract spatial correlation features of each load unit based on multi-dimensional feature data, and inputs the extracted spatial correlation features into the long short-term memory neural network model. The long short-term memory neural network model is used to extract time-series features representing the time-varying patterns of each load unit based on the received spatial correlation features, and to generate load forecast data for each load unit at multiple time scales in the future based on the extracted time-series features.

[0041] Specifically, in this application, the load forecasting model is a hybrid AI model that integrates a spatiotemporal graph neural network (STGNN) and a long short-term memory network (LSTM) to achieve accurate load forecasting at different time scales.

[0042] Specifically, such as Figure 2 As shown, the load forecasting model of this application adopts a serial structure. The input data layer inputs a 64-dimensional feature vector, the front stage is an STGNN spatial feature extraction layer, the back stage is an LSTM temporal feature modeling layer, and the output stage is a multi-scale forecasting head. The parameters and functional design of each layer of the load forecasting model of this application are as follows: Input layer: Receives a 64-dimensional feature vector after spatiotemporal fusion. The feature vector dimension consists of "static parameters + real-time operation + external environment + user behavior", ensuring that the model can make predictions based on sufficient historical spatiotemporal information.

[0043] STGNN Spatial Feature Extraction Layer: Its core function is to capture the spatial coupling relationships between different regions and loads, such as the impact of elevator heat dissipation on air conditioning load, and the matching relationship between lighting area distribution and personnel density. This layer includes graph construction units, spatial attention sublayers, and graph convolution sublayers, specifically: Graph Construction Unit: Used to construct a graph structure based on the spatial topology of different load units in the target building. A spatial association graph structure G=(V, E) is constructed using both physical and data association criteria. Node V represents each load unit within the building complex (e.g., air conditioning units on the 1st floor, lighting circuits on the 2nd floor, elevator groups on the 3rd floor, etc.; for example, it is divided into 32 nodes). The attribute of each node is the 64-dimensional feature vector of the corresponding load unit. The weight of edge E is calculated by weighting the physical distance (negative correlation) between two nodes with the historical load correlation coefficient (positive correlation). The weight formula is shown below:

[0044] in, For nodes With nodes The edge weights between the nodes are used to reflect the spatial association strength between the two nodes; For nodes and The load correlation coefficient; The physical distance between the two nodes; This is the maximum distance threshold (for example, it can be set to 50m). This is the weighting coefficient (which can be set to 0.7 to highlight the dominant role of data association).

[0045] Spatial Attention Sublayer: Multiple spatial attention heads are set, preferably five in this application. Each attention head maps node features to a 32-dimensional space through a linear transformation. The attention weight matrix is ​​calculated to highlight the feature contributions of highly correlated nodes (such as lighting and air conditioning nodes in densely populated areas). In this application, the attention weights are calculated using a scaled dot product mechanism, and the calculation formula is as follows:

[0046] in, For nodes For nodes Attention weights reflect the node's attention weights. Feature pairs of nodes The importance of feature extraction; , For nodes , eigenvectors; Let represent the feature vector of the k-th node. The feature dimension (e.g., K=32) represents the length of the feature vector. For nodes The set of neighboring nodes, where T represents the transpose operation.

[0047] Graph convolutional sublayer: This sublayer combines spatial attention-weighted node features with the spatial relational graph, reducing computational complexity and outputting a 32-dimensional spatial feature vector. The activation function of this sublayer uses... (Slope 0.01), used to alleviate the gradient vanishing problem.

[0048] The LSTM temporal feature modeling layer inherits the spatial features output by the STGNN and captures long-term temporal trends of the load, such as weekday / weekend load differences, peak-valley load fluctuations, and load changes due to seasonal changes. The LSTM temporal feature modeling layer in this application consists of a three-layer stacked LSTM structure, with each LSTM layer containing 128 hidden neurons. Interlayer connections are... Regularization ( To prevent overfitting, a gating rate of 0.2 is used. The specific gating mechanism is designed as follows: Forgotten Gate: adopts The activation function takes the spatial features of the current time step and the hidden state of the previous time step as input, and outputs a forgetting weight vector that determines the proportion of historical temporal features to be retained. The specific formula is shown below:

[0049] in, for The output weight vector of the time-forget gate (value range [0,1]), where 0 indicates that historical information is completely forgotten and 1 indicates that historical information is completely retained; for Activation function; The forget gate weight matrix (dimension 128×(128+32)). The hidden state of the previous time step. For the spatial features of the current time step, This is the bias vector.

[0050] Input gate: by Activated update weight and The activated candidate features are used to filter and update the key temporal features at the current time step, as shown below:

[0051] in, for The update weight vector of the input gate at time step (value range [0,1]), where 0 indicates that the corresponding feature is not updated and 1 indicates that the corresponding feature is fully updated; Here is the weight matrix of the input gate; The bias vector for the input gate; for The candidate cell state at each time step stores new feature information for the current time step; is the hyperbolic tangent activation function (value range [-1, 1]); This is the weight matrix for the candidate cell states; is the bias vector for the candidate cell state.

[0052] Cell state update: Combining the results of the forget gate and the input gate, the cell state is updated to store long-term time information, as shown below:

[0053] in, for The cell state is updated in real time and stored up to the end. The long-period time characteristics of a moment; for Cellular state at any given moment; This is element-wise multiplication (Hadamard product).

[0054] Output gate: Filters key information from the cell state as the hidden state at the current time step, outputting a 64-dimensional time-space fusion feature, as shown below:

[0055] in, for The output gate's filtering weight vector (value range [0,1]) is set at each time step. 0 indicates that the corresponding cell state information is masked, and 1 indicates that the corresponding cell state information is output. This is the weight matrix of the output gate; This is the bias vector for the output gate; for The output hidden state at each time step is a 64-dimensional time-space fusion feature.

[0056] Multi-scale prediction: This application employs a parallel fully connected layer structure, setting different prediction branches for three scales: short-term (15min-2h), medium-term (2h-24h), and long-term (24h-7d). Each branch achieves scale adaptation by sampling at different time steps and matching the output dimension. Short-term forecast branch: Input the hidden state of the last 4 time steps of the LSTM layer (corresponding to 40 minutes of historical data), pass through a fully connected layer (64→32 dimensions) and ReLU activation, and output the load forecast value for the next 12 time steps (1 point every 15 minutes, for a total of 2 hours); Mid-term forecast branch: Input the hidden state of the last 12 time steps of the LSTM layer (corresponding to 2 hours of historical data), after passing through 2 fully connected layers (64→48→24 dimensions) and ReLU activation, output the load forecast value for the next 24 time steps (1 point per hour, for a total of 24 hours); Long-term forecast branch: Input the hidden states of all 24 time steps of the LSTM layer (corresponding to 4 hours of historical data), pass through 2 fully connected layers (64→64→7 dimensions) and ReLU activation, and output the load forecast values ​​for the next 7 time steps (1 point every 24 hours, for a total of 7 days); Each branch output layer uses a linear activation function to ensure that the predicted values ​​are consistent with the actual load dimensions.

[0057] In step S200, the training process of the load prediction model includes: S210. Obtain historical multidimensional feature data of other buildings, and train the load forecasting model at each time scale using the historical multidimensional feature data of other buildings to determine the model parameters of the load forecasting model. S220. Obtain historical multidimensional feature data of the target building, and train the load forecasting model at various time scales using the historical multidimensional feature data of the target building to update the model parameters of the load forecasting model; wherein, when training the load forecasting model at different time scales, the weights of static attribute parameters, dynamic operating parameters, external environment parameters or user behavior parameters in the input historical multidimensional feature data are dynamically adjusted based on the current training time scale.

[0058] To ensure the prediction accuracy and generalization ability of the model in dynamic building cluster scenarios, this application adopts a three-level training strategy of "pre-training-fine-tuning-online update", combined with adaptive optimization algorithm and regularization for model training. The specific training process is as follows: Dataset Construction and Preprocessing: 12 months of historical data on the target building complex were collected, covering different seasons, weekdays / weekends, and holidays, with a total sample size of over 100,000. This data was divided into training, validation, and test sets in an 8:1:1 ratio. In this application, in addition to routine data cleaning and normalization, the data preprocessing also included adding ±5% Gaussian noise to the time series data and shifting the time step (±1 sampling period) to generate enhanced samples, thereby improving the model's generalization ability.

[0059] Pre-training phase: A transfer learning approach is adopted, using a publicly available building load dataset to pre-train the model to initialize the parameter weights of STGNN and LSTM, shortening the training cycle of the model in the target scenario. Pre-training can use a fixed learning rate (1e-3), 50 training epochs, and mean squared error (MSE) as the loss function. Pre-training stops when the validation set loss does not decrease for 5 consecutive epochs.

[0060] Fine-tuning phase: The pre-trained model is transferred to the target scene dataset, i.e., historical data of the target buildings. An adaptive learning rate algorithm (AdamW optimizer) is used to dynamically adjust the learning rate. The initial learning rate can be set to 1e-3. The learning rate is maintained as the validation set loss decreases. If the loss does not decrease for three consecutive rounds, the learning rate is decayed to 1 / 10 of its original value, with a minimum learning rate of 1e-5. A hybrid loss function is also introduced, combining MSE and Mean Absolute Percentage Error (MAPE) to balance prediction errors at different load levels. The hybrid loss function is shown below:

[0061] in, This is the weighting coefficient (set to 0.6). For predicted values, The values ​​represent the true values. The fine-tuning phase consists of 30 training epochs, employing an early stopping strategy (patience=5) to prevent overfitting.

[0062] Regularization and optimization: In addition to dropout regularization in the LSTM layer, this application also introduces L2 regularization (weight decay coefficient 1e-4) in the fully connected layer and uses gradient pruning technique (gradient norm threshold 5.0) to prevent gradient explosion; during training, batch normalization is used to accelerate model convergence and improve training stability.

[0063] To ensure the accuracy of predictions at each scale, this application dynamically adjusts the weights of each feature by constructing a scale-adaptive feature focusing mechanism. In this application, the multiple time scales include at least a first time scale, a second time scale, and a third time scale; wherein the duration of the first time scale is shorter than that of the second time scale, and the duration of the second time scale is shorter than that of the third time scale. The weights of static attribute parameters, dynamic operating parameters, external environment parameters, or user behavior parameters in the input historical multidimensional feature data are dynamically adjusted based on the current training time scale, including: if the current training time scale is the first time scale, increasing the weights of the dynamic operating parameters and user behavior parameters in the input historical multidimensional feature data; if the current training time scale is the second time scale, increasing the weights of the dynamic operating parameters and external environment parameters in the input historical multidimensional feature data, with the weights of the dynamic operating parameters and external environment parameters being equal; if the current training time scale is the third time scale, increasing the weights of the external environment parameters in the input historical multidimensional feature data.

[0064] Specifically, the scale-adaptive feature focusing mechanism dynamically adjusts feature weights at different scales through an attention mechanism. Short-term forecasting (15min-2h) focuses on real-time operational features (such as current air conditioner power and elevator operating status) and user behavior features (such as real-time pedestrian density), with a feature weight ratio of 60%. Medium-term forecasting (2h-24h) balances real-time features and external environmental features (such as hourly weather forecasts and electricity price periods), with each having a weight ratio of 40%. Long-term forecasting (24h-7d) focuses on external environmental features (such as daytime maximum temperature and holiday arrangements) and historical load trend features, with a weight ratio of 70%.

[0065] This application also employs a multi-scale output fusion strategy to dynamically fuse the output results. Specifically, the output includes load forecast data for at least one load unit of the target building at multiple time scales. This includes: if the deviation between the load forecast data of at least one load unit of the target building at the first time scale and its load forecast data at the corresponding time period at the second time scale is greater than a preset first deviation threshold, updating the load forecast data of at least one load unit of the target building at the first time scale with the load forecast data at the corresponding time period at the second time scale; and if the deviation between the load forecast data of at least one load unit of the target building at the second time scale and its load forecast data at the corresponding time period at the third time scale is greater than a preset second deviation threshold, updating the load forecast data of at least one load unit of the target building at the second time scale with the load forecast data at the corresponding time period at the third time scale.

[0066] Specifically, the multi-scale output fusion strategy employs a hierarchical output approach of "coarse prediction - fine correction." Long-term forecasts serve as prior information for medium-term forecasts, while medium-term forecasts act as constraints for short-term forecasts. A gating fusion unit integrates forecasts from different scales to correct prediction biases at a single scale. For example, in short-term forecasting, if the predicted value deviates more than 10% from the corresponding trend of the medium-term forecast, the gating unit adjusts the short-term forecast value to ensure consistency across time scales.

[0067] Furthermore, this application further improves the prediction accuracy through a real-time error correction mechanism. Therefore, in this application, outputting load prediction data for at least one load unit of the target building at multiple time scales in the future includes: for each time scale: acquiring the load prediction data and the measured load data of the previous time; determining the correction coefficient of the load prediction data at the current time based on the error of the load prediction data and the measured load data at the previous time; and correcting the load prediction data at the current time with the correction coefficient.

[0068] Specifically, the real-time error correction mechanism introduces a sliding window correction module to compare hourly forecasts with actual load data, calculate error correction coefficients, and adjust subsequent forecasts based on the correction coefficients when the forecast error at a certain scale exceeds a threshold (5% for short-term, 8% for medium-term, and 12% for long-term). The correction coefficients for the current load forecast data are determined based on the error between the previous load forecast data and the previous measured load data, including: The correction factor is calculated using the Exponential Moving Average (EMA) and is determined by the following formula:

[0069] in, Let be the correction factor at time t. The correction factor at time t-1 This is the smoothing coefficient (set to 0.8). The load measurement data is at time t-1. The load forecast data at time t-1 is corrected to the following forecast value: .

[0070] Accuracy verification and results of the model in this application: In the target commercial complex scenario, the model was continuously tested for one month. The prediction accuracy indicators at each scale are as follows: short-term prediction (15min-2h) MSE is 0.0025, MAPE is 4.2%; medium-term prediction (2h-24h) MSE is 0.0068, MAPE is 7.5%; long-term prediction (24h-7d) MSE is 0.015, MAPE is 11.3%, all of which meet the preset accuracy requirements (short-term ≤5%, medium-term ≤8%, long-term ≤12%), which is more than 40% higher than the accuracy of the single LSTM model (short-term MAPE 10.8%, medium-term 16.2%, long-term 22.5%).

[0071] This application establishes an optimization system encompassing four major objectives: energy efficiency, economy, comfort, and grid response. It employs the Analytic Hierarchy Process (AHP) to determine the objective weights, thus avoiding the limitations of single-objective optimization. Specifically, it includes: Core objectives defined as follows: Energy efficiency objective: Minimize total energy consumption of air conditioning, lighting, and elevators, with energy consumption per unit area reduced by ≥20% compared to the benchmark value; Economic objective: Minimize electricity purchase costs, with peak load reduction rate ≥15% and peak energy storage charging capacity maximized; Comfort objective: Maintain indoor temperature at 22-26℃, lighting brightness at 300-500 lux (office area) / 500-700 lux (commercial area), average elevator waiting time ≤30s, and comfort compliance rate ≥95%; Grid response objective: Green electricity absorption rate ≥80%, and grid voltage fluctuation controlled within ±5%.

[0072] Target weight allocation: Using the AHP method and combined with energy management experience scoring, the weights were determined as follows: energy efficiency target 0.35, economic target 0.3, comfort target 0.25, and grid response target 0.1. The weight consistency test CR value was ≤0.1 to ensure the rationality of the weight allocation.

[0073] Optimization function construction: Construct a multi-objective optimization function, where the objective function is specifically:

[0074] in, Let be the energy consumption function. For cost function, For comfort function, Let be the power grid response function. The weights of the energy consumption function, The weights of the cost function, The weights of the comfort function, The weights represent the weights of the grid response function. The energy consumption function, cost function, comfort function, and grid response function represent the mapping relationship between the target building's energy consumption, cost, comfort value, and grid response value and the control parameters of each load unit, respectively. The control parameters of each load unit are the control parameters of each power-consuming device in the target building.

[0075] Wherein, f1 is the energy efficiency objective function (unit: kWh), which can be used to represent the total energy consumption of air conditioning, lighting, and elevators; the smaller its value, the higher the energy efficiency. f2 is the economic objective function (unit: yuan), which is used to represent the total cost of electricity purchase for the building complex; the smaller its value, the better the economic efficiency. f3 is the comfort objective function, whose value range can be [0,1], used to represent the comfort compliance rate; the closer its value is to 1, the higher the comfort level. Therefore, (1-f3) can be used to convert it into a minimization objective. f4 is the grid response objective function, whose value range can be [0,1], used to represent the comprehensive score of green electricity absorption rate and voltage stability; the closer its value is to 1, the better the grid response effect. Therefore, (1-f4) can be used to convert it into a minimization objective. It is understood that the specific forms of objective functions f1-f4 can be adjusted or determined according to needs, and this application does not limit the specific forms of f1-f4.

[0076] For example, f1 = (E_ac + E_light + E_elev) / base day energy consumption, where E_ac is the total energy consumption of the air conditioning system in 24 hours, E_light is the total energy consumption of the lighting system in 24 hours, and E_elev is the total energy consumption of the elevator system in 24 hours; f2 = Σ[electricity price (t) × purchased power (t)] / base day daily cost; f3 = (temperature compliance time / 24) × a + (brightness compliance time / 24) × b + (elevator waiting time compliance time / 24) × c, where a, b, and c are weights; f4 = (photovoltaic absorption / photovoltaic reference capacity) × α + (1 - voltage fluctuation rate) × β, where α and β are weights. Constraints can be set as follows: 22℃ ≤ T_ac ≤ 26℃, 300lux ≤ L_light ≤ 700lux, air conditioning power ≤ 500kW, simultaneous elevator operation ≤ 8 units, etc. It is understood that the objective function and constraints can be set according to specific needs, and are not limited here.

[0077] Based on multi-scale prediction results and multi-objective optimization functions, this application uses an improved non-dominated sorting genetic algorithm (NSGA-Ⅲ) to generate coordinated control strategies for air conditioning, lighting, and elevators. In step S300, the pre-constructed objective function is optimized and solved using an improved objective optimization algorithm based on the coupling relationship of various power consumption devices in the target building, including: S1. Construct a coupling factor representing the coupling relationship of each power-consuming device, wherein the coupling factor is used to represent the degree of influence of the operation of any power-consuming device on the control parameters of other power-consuming devices. Based on the traditional NSGA-Ⅲ, this application introduces load coupling factor encoding, embedding the coupling relationship of air conditioning, lighting, and elevator into chromosome encoding (encoding length is 32 bits).

[0078] S2. Initialize the population by encoding the control parameters and coupling factors of each load unit into an individual, with each individual representing a solution to the objective function. For example, each chromosome is 32 bits long and is divided into air conditioning, lighting, and elevator gene segments. Coupling factors representing the sensitivity coefficient of air conditioning to elevator heat dissipation and the compensation coefficient of lighting color temperature to air conditioning set temperature are embedded in the chromosomes. Specifically, the air conditioning gene segment represents the air conditioning control strategy, such as set temperature and fan speed; the lighting gene segment represents the lighting control strategy, such as brightness and color temperature; and the elevator gene segment represents the elevator control strategy, such as the number of elevators in operation, operating speed, and grouping strategy.

[0079] Specifically, this application divides the 32-bit chromosome into multiple fields, each corresponding to the control parameters and coupling factors of a type of load (air conditioning, lighting, elevator). For example, in a specific instance, the chromosome structure can be represented as follows: Air conditioning: bits 0-7, air conditioning start / stop / operating power / set temperature (discrete encoding, such as 00101011 representing a set temperature of 27℃); Lighting: bits 8-15, lighting switch / brightness level / zone control (such as 11000101 representing 70% brightness in zone 3); Elevator: bits 16-23, elevator start / stop / operating mode / floor priority (such as 01101001 representing peak mode + priority service for floors 10-20); Coupling factor: bits 24-31, characterizing the coupling strength between the three types of loads, for example: bits 24-26: Air conditioning-lighting coupling factor: Value range: 000(0)~111(7), the larger the value, the stronger the coupling (e.g., 101(5) means that for every 1℃ drop in air conditioning temperature, the lighting brightness needs to be reduced by 10%); 27-29: Air conditioning-elevator coupling factor: Value range: 000(0)~111(7), e.g., 011(3) means that when the air conditioning load exceeds the threshold, the elevator start-stop interval is extended by 5s; 30-31: Lighting-elevator coupling factor: Value range: 00(0)~11(3), e.g., 10(2) means that when all lighting is on, the elevator prioritizes serving lower floors (reducing energy consumption of higher floors). The specific structural composition of the chromosome can be adjusted according to the actual situation, and is not limited here.

[0080] Meanwhile, this application employs an adaptive crossover operator (crossover probability 0.7-0.9) and a mutation operator (mutation probability 0.01-0.03) to enhance the algorithm's global search capability; and sets the population size to 200 and the number of iterations to 50 to ensure efficient policy generation (single policy generation time ≤ 30s).

[0081] S3. For each individual in the population, calculate the fitness value of the current individual using the objective function.

[0082] Suppose that in a chromosome, the air conditioning segment gene encodes a set temperature of 27.5℃ and a fan speed of 75% of maximum speed. After correction using the thermal coupling factor, the final set temperature is corrected to 26.8℃. Then, based on the final set temperature and fan speed, the daily energy consumption of the air conditioning system under the current strategy is calculated. Similarly, the daily energy consumption of the lighting system and elevator system of the current chromosome is calculated to obtain the value of f1. By analogy, based on the obtained energy consumption, the values ​​of f2-f4 can be calculated respectively, thus obtaining the fitness value of the current individual.

[0083] S4. Determine the non-dominance relationship of each individual based on their fitness values, and stratify and rank them according to their non-dominance relationships. First, calculate the dominance relationship of each individual. For example, smaller values ​​for f1 and f2 are better, while larger values ​​for f3 and f4 are better. Assume individual 1 has f=(0.85,0.78,0.88,0.72) and individual 2 has f=(0.82,0.85,0.85,0.75). Then individual 1 is poor in f1 (0.85>0.82), good in f2 (0.78<0.85), good in f3 (0.88>0.85), and good in f4 (0.72<0.75). Therefore, its dominance relationship cannot be determined, and its dominance relationship is non-dominance. This process continues, resulting in multiple strata based on non-dominance relationships. The first stratum consists of individuals not dominated by any other individuals, the second stratum consists of individuals dominated by the first stratum but not dominated by other individuals, and so on. The specific process of performing non-dominated sorting on each individual entity in the construction is existing technology and is not limited here.

[0084] S5. Based on the hierarchical ranking results of each individual, determine the parent individuals. Perform crossover and mutation operations on the parent individuals to update the individuals in the current population. If the convergence condition is met, the individual with the best fitness value in the current population is taken as the optimal solution of the objective function, and the optimal solution is used as at least one load unit, i.e., the coordinated control strategy of various power consumption devices such as lighting, air conditioning, and elevators in the target building; otherwise, return to step S3. For example, randomly select two individuals A and B from the current population. Compare the non-dominated levels of individuals A and B according to the hierarchical ranking results of each individual. Place the individual with the better non-dominated level into the parent pool until the number of individuals in the parent pool reaches the preset population size. Randomly pair the individuals in the parent pool and perform crossover and mutation operations according to the crossover and mutation probabilities to update the individuals in the current population. It can be understood that the convergence condition can be reaching the maximum number of iterations, or the Pareto front being stable, and the change of the non-dominated solution set being very small for several consecutive generations, etc.

[0085] Once the convergence condition is met, determine the individual with the optimal fitness value in the current population as the optimal solution for the objective function, or calculate the TOPSIS value of each individual in the current population and select the individual with the optimal TOPSIS value as the optimal solution. The detailed control strategies of this application include: Air conditioning control: Based on the predicted load and indoor temperature, dynamically adjust the set temperature (increase by 1-2℃ during peak hours and decrease by 1-2℃ during off-peak hours), fan speed (adjust to 30%-100% according to the flow density), and start / stop sequence (rotate shutdown in different areas during non-working hours); Lighting control: Adopt a zoned dimming strategy, maintain standard brightness in densely populated areas, reduce brightness to below 50 lux in unoccupied areas, dynamically adjust the brightness in meeting rooms according to the number of people and natural light intensity, and simultaneously optimize the air conditioning load through color temperature adjustment (5000-6500K cool light during peak hours and 3000-4000K warm light during off-peak hours); Elevator control: During peak hours, adopt zoned and grouped operation (e.g., 1-5 floors, 6-10 floors grouped) to reduce cross-zone stops; During off-peak hours, adopt the "nearest stop + scheduled dispatch" mode; During low-peak hours, reduce the number of elevators in operation (retain 30%-50% of the number of elevators), and align the elevator start / stop times with the low-load periods of the air conditioning system.

[0086] Specifically, the detailed load-sharing control strategy of this application includes: Air conditioning control strategy: Based on predicted load and coupled correction values, dynamic adjustments are made. Core parameters include set temperature, fan speed, and start / stop sequence. Set temperature = base predicted temperature + elevator heat dissipation correction value + lighting heat dissipation correction value. Base predicted temperature is calculated based on outdoor temperature and pedestrian density (for every 1°C increase in outdoor temperature, the base temperature increases by 0.3°C; for every 0.1 people / m² increase in pedestrian density, the base temperature decreases by 0.2°C). Fan speed = base speed × (actual load / predicted load) × pedestrian density coefficient (1.1 for densely populated areas, 0.9 for sparsely populated areas). Start / stop sequence adopts a zoned rotational shutdown approach, prioritizing the shutdown of air conditioning in unoccupied and low-load areas, with a rotational shutdown interval of ≥10 minutes to avoid sudden load changes.

[0087] Lighting control strategy: Adopting a "zoning-time-color temperature linkage" mode, wherein: Zoning brightness: Adjusted in real time according to the density of people (office area: 300-500 lux when people flow ≥ 0.5 people / ㎡, 100-300 lux when people flow < 0.5 people / ㎡; commercial area: 500-700 lux when people flow ≥ 1 person / ㎡, 300-500 lux when people flow < 1 person / ㎡); Time-sharing strategy: Different brightness curves are preset for weekdays and holidays, with brightness reduced by 10% during peak hours (9:00-15:00) and only corridor lighting retained during off-peak hours (22:00-7:00); Color temperature linkage: Linked with the air conditioning temperature setting, with a color temperature of 5000-6500K (cool light) when the temperature is ≥ 25℃ and a color temperature of 3000-4000K (warm light) when the temperature is ≤ 23℃, using psychological effects to help improve comfort.

[0088] Elevator control strategy: Adopting a "time-segmentation-grouping-reservation linkage" group control mode, with the following elevator operation rates: 80%-100% during peak hours (7:30-9:30, 17:30-19:30), 50%-70% during off-peak hours (9:30-17:30), and 30%-50% during low-peak hours (19:30-7:30); Grouping strategy: High zones (6-10 floors) and low zones (1-5 floors) are operated in groups, and a "shuttle elevator" is activated when calling across zones to reduce unnecessary stops; Reservation linkage: Connecting to the conference reservation system, the elevator is automatically dispatched to the conference floor 5 minutes before the end of a large conference, with an average waiting time of ≤30 seconds.

[0089] Understandably, once the global coordinated control strategy for at least one load unit—that is, each power-consuming device in the target building, such as lighting, air conditioning, and elevators—is obtained, each load unit can dynamically adjust the global coordinated control strategy based on factors such as its location, equipment characteristics, real-time status, and user feedback. For example, taking load unit 1 as the air conditioning unit in the lobby on the first floor, with the received global control strategy based on a set temperature of 27℃, the static attribute parameters include: lobby on the first floor (west-facing, large glass curtain wall); user behavior parameters include: real-time pedestrian density of 0.3 people / ㎡ (low); external environmental parameters include: outdoor temperature of 33℃, strong solar radiation; historical data information includes: afternoon cooling load is 15% higher than predicted due to western exposure. The global control strategy is then modified according to preset rules. For example, due to western exposure and high temperature, the temperature is automatically reduced by 0.5℃, that is, the actual set temperature is 26.5℃. Due to low population density, the fan speed is reduced to 70% (energy saving). The coupling factor between air conditioning and lighting is found to be 5. If the lighting brightness is high, the temperature can be appropriately increased. For example, if the current lighting brightness on the first floor is only 50%, compensation is not triggered, and the temperature is maintained at 26.5℃.

[0090] Furthermore, after generating specific control strategies for air conditioning, lighting, and elevators through optimal decoding, this application also ensures conflict-free and optimal strategy execution by constructing a coupled conflict resolution knowledge base and strategy priority rules. Specifically, the collaborative control strategy is dynamically adjusted based on the real-time operating status of at least one load unit, including: real-time acquisition of dynamic operating parameters, external environmental parameters, and user behavior parameters of the corresponding areas of each load unit in the target building; if the conflict resolution knowledge base determines that at least one of the dynamic operating parameters, external environmental parameters, and user behavior parameters of at least one load unit meets the conflict condition, the conflict resolution knowledge base determines the corresponding conflict resolution rule, and the current collaborative control strategy is adjusted according to the conflict resolution rule; the conflict resolution knowledge base includes at least conflict resolution rules corresponding to different conflict conditions, and the conflict conditions include at least one of the dynamic operating parameters, external environmental parameters, and user behavior parameters of at least one load unit meeting the preset conditions.

[0091] The coupling conflict resolution mechanism of this application includes: A conflict resolution knowledge base is constructed, specifically containing resolution rules for three core conflict scenarios. The triggering conditions for these rules are based on real-time load deviation and comfort feedback, as follows: Scenario 1: Conflict between peak air conditioning load and insufficient lighting: Triggering conditions are that the actual air conditioning load is 15% higher than the predicted load and there are ≥3 complaints about lighting brightness per hour. Resolution rules: Air conditioning side: Increase the set temperature by 0.5-1℃, keep the fan speed unchanged (avoid a sudden drop in comfort); Lighting side: Increase the color temperature by 500K, keep the brightness at the current value (cool light improves perceived visual brightness); Elevator side: Reduce the number of elevators operating across zones by 10%-15%, reducing the impact of elevator heat dissipation on the air conditioning. This rule can reduce the air conditioning load by 8%-10%, while maintaining a comfort score above 4 out of 5.

[0092] Scenario 2: Long elevator waiting time + low air conditioning load conflict: Triggering conditions are an average elevator waiting time > 35 seconds and an air conditioning load < 20% of the predicted load. Resolution rules: Elevator side: Increase the number of elevators in operation by 10%, adopting a "nearest stop" priority mode; Air conditioning side: Utilize the increased load from elevator heat dissipation to appropriately lower the set temperature by 0.3-0.5℃ to improve comfort; Lighting side: Increase the brightness of the elevator waiting area by 50 lux to alleviate waiting anxiety. This rule can shorten the elevator waiting time to 25-30 seconds and increase the air conditioning load utilization rate to over 90%.

[0093] Scenario 3: Sudden Increase in Green Energy Output + Conflict in Load Absorption: Triggering conditions are that actual photovoltaic output > predicted output by 30% and curtailment rate > 10%. Resolution Rules: Air Conditioning Side: Turn on air conditioners in low-load areas in advance, lowering the set temperature by 1°C to utilize the air conditioner's cold storage characteristics to absorb green energy; Lighting Side: Turn on decorative lighting in commercial areas (brightness 300-400 lux, without affecting main lighting); Elevator Side: Conduct elevator no-load test runs during off-peak hours (once per hour, 5 minutes each time), while adjusting the charging time of charging equipment (such as elevator backup batteries) to the peak green energy output period. This rule can increase the green energy absorption rate to over 90% and reduce the curtailment rate to below 5%.

[0094] Meanwhile, this application adopts an architecture of "edge computing + cloud backup" to achieve real-time execution and dynamic optimization of strategies, ensuring control response speed: The real-time execution channel includes: control strategies are sent to each load controller through an industrial-grade edge gateway (supporting offline operation). The air conditioning controller uses a PID algorithm (control accuracy ±0.2℃), the lighting controller uses a PWM dimming algorithm (dimming accuracy ±10lux), and the elevator controller uses a group control algorithm. The instruction issuance delay is ≤1s, and the execution response time is ≤3s.

[0095] 1) Air Conditioner Controller PID Algorithm Design and Accuracy Assurance: An incremental PID control algorithm is adopted. Considering the multi-zone collaborative characteristics of central air conditioning, a "master controller-zone slave controller" architecture is constructed. The master controller receives gateway commands and distributes them to each zone slave controller. PID parameters are dynamically configured: the proportional coefficient Kp is set to 3.5 (adaptively adjusted to ±0.5 during load fluctuations), the integral coefficient Ki is set to 0.08 (to eliminate static errors), and the derivative coefficient Kd is set to 0.2 (to suppress overshoot). The control flow is as follows: real-time acquisition of indoor temperature (sampling frequency 10Hz) and the deviation from the set temperature → calculation of fan speed and electronic expansion valve opening adjustment via PID algorithm → output of 4-20mA analog signal to the actuator. To ensure a control accuracy of ±0.2℃, a temperature compensation mechanism is introduced: when the outdoor temperature fluctuation exceeds 5℃ or the change in pedestrian density is ≥0.3 people / ㎡, the PID parameter deviation value is automatically corrected (correction range ±0.1℃), and an anti-interference filtering algorithm is used to eliminate high-frequency noise from the sensors.

[0096] 2) PWM Dimming Algorithm Implementation for Lighting Controller: An intelligent lighting controller based on the ARM Cortex-M4 core is used. Precise brightness adjustment is achieved through PWM (Pulse Width Modulation) technology, supporting a dimming range of 0-100%. The core PWM parameters are configured as follows: carrier frequency 20kHz (to avoid visible flicker), 16-bit resolution (corresponding to 65536 levels of dimming accuracy), and linear brightness output is achieved by adjusting the duty cycle of the high-level pulse. Preset dimming curves are provided for different lighting scenarios (office areas, commercial areas, corridors): Office areas use a "slow start / stop" curve (the start / stop process lasts 2 seconds to avoid sudden brightness changes), while commercial areas use a "stepped dimming" curve (adjusted in 50-lux increments for smoother response). ±10lux accuracy is ensured through closed-loop feedback: each intelligent lamp has a built-in light sensor (accuracy ±1lux, sampling frequency 5Hz) that provides real-time feedback of the actual brightness to the controller. When the deviation from the target brightness exceeds 5lux, a fine-tuning of the PWM duty cycle is triggered (adjustment step size 0.1%).

[0097] 3) Elevator Group Control Algorithm Logic and Execution Mechanism: The system employs a "prediction-based dynamic grouping + nearest-neighbor" group control algorithm. The controller uses the LCC-3000 dedicated elevator group control module, supporting coordinated scheduling of 12 elevators. The core algorithm logic is as follows: It receives the number of operating elevators and grouping strategy instructions from the gateway → combines real-time collected elevator position, car load, and call signal data → dynamically divides the system into high-zone (6-10 floors) and low-zone (1-5 floors) operating groups, automatically assigning a "shuttle elevator" when calling across zones. Execution response process: After a call signal is triggered, the controller completes optimal elevator matching within 50ms and sends operating instructions (such as car door opening / closing and speed adjustment) via the CANopen bus. The elevator start-up response time is ≤2s. Throughout the entire operation, speed closed-loop control (sampling frequency 20Hz) ensures a stopping accuracy of ±5mm.

[0098] Feedback data acquisition: Real-time acquisition of load operation parameters (such as actual air conditioning power and lighting brightness), environmental parameters (indoor temperature and humidity), and user comfort feedback after regulation. The sampling frequency is 30 seconds / time to form a feedback dataset.

[0099] Closed-loop optimization mechanism: The feedback dataset is input into the AI ​​prediction model, and the model parameters are updated using the stochastic gradient descent (SGD) incremental learning algorithm (each update time ≤ 30s); at the same time, the deviation between the actual effect and the optimization target is calculated. If the deviation exceeds 5% (such as the energy consumption reduction not reaching the target), the strategy is triggered for secondary optimization to ensure that the regulation effect is stable and meets the target.

[0100] In summary, this application constructs a four-dimensional feature system based on multi-dimensional feature fusion for accurate prediction. It incorporates user behavior features such as meeting reservations, comfort feedback, and device coupling into the prediction, and combines this with an STGNN-LSTM hybrid model to capture spatiotemporal correlations, resulting in a short-term prediction error of ≤5%, which is more than 40% more accurate than traditional models. Furthermore, it constructs a load-coupled driven collaborative control logic, introducing load coupling factors and a conflict resolution knowledge base to resolve conflicts arising from independent control of air conditioning, lighting, and elevators. For example, color temperature adjustment assists air conditioning in energy saving, increasing the overall energy consumption reduction rate by 8%-12% compared to independent control. Finally, it establishes a closed-loop optimization mechanism for incremental learning, updating the model online through real-time feedback data to adapt to dynamic scenarios such as seasonal changes and personnel movement, ensuring long-term accuracy decay of ≤3%, thus solving the problems of static training and dynamic failure in traditional models.

[0101] like Figure 3 As shown, in a second aspect, this application provides an artificial intelligence-based flexible load coordination and control device for building complexes, comprising: The data acquisition module is configured to acquire multidimensional feature data of the target building, including static attribute parameters, dynamic operating parameters, external environment parameters, and user behavior parameters of the target building. The load forecasting module is configured to take multidimensional feature data as input and output load forecasting data for at least one load unit of the target building in the future at multiple time scales after a pre-trained load forecasting model. The coordinated control strategy module is configured to optimize and solve a pre-built objective function based on load forecast data at multiple time scales through an improved objective optimization algorithm based on the coupling relationship of each power consumption device in the target building, thereby obtaining a coordinated control strategy for at least one load unit. The coordinated control strategy includes the control parameters of the corresponding load unit. The coordinated control module is configured to execute a coordinated control strategy on at least one load unit and dynamically adjust the coordinated control strategy according to the real-time operating status of at least one load unit.

[0102] It is understood that those skilled in the art will clearly recognize that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0103] In a third aspect, this application provides a machine-readable storage medium storing instructions that, when executed by a processor, cause the processor to perform the artificial intelligence-based flexible load coordination control method for building clusters as described above.

[0104] In a fourth aspect, this application provides a terminal device including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the artificial intelligence-based flexible load collaborative control method for building clusters as described above.

[0105] like Figure 4 The diagram shown is a schematic representation of a terminal device provided in an embodiment of this application. Figure 4As shown, the terminal device 10 of this embodiment includes a processor 100, a memory 101, and a computer program 102 stored in the memory 101 and executable on the processor 100. When the processor 100 executes the computer program 102, it implements the steps in the above method embodiments. Alternatively, when the processor 100 executes the computer program 102, it implements the functions of each module / unit in the above device embodiments.

[0106] For example, computer program 102 may be divided into one or more modules / units, one or more of which are stored in memory 101 and executed by processor 100 to complete this application. One or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 102 in terminal device 10.

[0107] Terminal device 10 may be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. Terminal device 10 may include, but is not limited to, a processor 100 and a memory 101. Those skilled in the art will understand that... Figure 4 This is merely an example of terminal device 10 and does not constitute a limitation on terminal device 10. It may include more or fewer components than shown, or combine certain components, or different components. For example, terminal device may also include input / output devices, network access devices, buses, etc.

[0108] Processor 100 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0109] The memory 101 can be an internal storage unit of the terminal device 10, such as a hard disk or RAM of the terminal device 10. The memory 101 can also be an external storage device of the terminal device 10, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or FlashCard equipped on the terminal device 10. Furthermore, the memory 101 can include both internal and external storage units of the terminal device 10. The memory 101 is used to store computer programs and other programs and data required by the terminal device 10. The memory 101 can also be used to temporarily store data that has been output or will be output.

[0110] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0111] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0112] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for coordinated control of flexible loads in building complexes based on artificial intelligence, characterized in that, include: Acquire multidimensional feature data of the target building, including static attribute parameters, dynamic operating parameters, external environment parameters, and user behavior parameters of the target building; Using the multidimensional feature data as input, the pre-trained load prediction model outputs load prediction data for at least one load unit of the target building in the future at multiple time scales. Based on the load forecast data at multiple time scales, the pre-constructed objective function is optimized and solved by an improved objective optimization algorithm based on the coupling relationship of each power consumption device in the target building, thereby obtaining the coordinated control strategy of at least one load unit, wherein the coordinated control strategy includes the control parameters of the corresponding load unit. The coordinated control strategy is executed on the at least one load unit, and the coordinated control strategy is dynamically adjusted according to the real-time operating status of the at least one load unit.

2. The method for coordinated control of flexible loads in building complexes based on artificial intelligence according to claim 1, characterized in that, The static attribute parameters include the building structure parameters of the target building, the rated parameters of each power consumption device, and building information model data. The building information model data includes spatial topology data representing the spatial relationship between different load units in the target building. The dynamic operating parameters include at least one of the following: real-time operating data of each power consumption device in the area corresponding to each load unit, temperature and humidity, CO2 concentration, light intensity, and output data of the distributed photovoltaic / energy storage equipment of the target building. The external environmental parameters include at least one of the following: meteorological data, electricity price data, and photovoltaic power output forecast data for future periods; The user behavior parameters include at least one of the following: the pedestrian density in the area corresponding to each load unit in the target building, the number of users making reservations in the area corresponding to each load unit, and user feedback on the comfort level of the area corresponding to each load unit.

3. The method for coordinated control of flexible loads in building complexes based on artificial intelligence according to claim 2, characterized in that, The load forecasting model includes: Serial spatiotemporal graph neural network model and time sequence neural network model; The spatiotemporal graph neural network model is constructed based on the spatial topology data of the target building, and the temporal neural network model is a long short-term memory neural network model. The spatiotemporal graph neural network model is used to extract the spatial correlation features of each load unit based on the multidimensional feature data, and the extracted spatial correlation features are input into the long short-term memory neural network model. The long short-term memory neural network model is used to extract time-series features representing the time-varying patterns of each load unit based on the received spatial correlation features, and to generate load prediction data for each load unit at multiple time scales in the future based on the extracted time-series features.

4. The method for coordinated control of flexible loads in building complexes based on artificial intelligence according to claim 1, characterized in that, The training process of the load prediction model includes: Obtain historical multidimensional feature data of other buildings, and train the load prediction model at each time scale using the historical multidimensional feature data of other buildings to determine the model parameters of the load prediction model; The historical multidimensional feature data of the target building is obtained, and the load prediction model is trained at each time scale using the historical multidimensional feature data of the target building to update the model parameters of the load prediction model. When training the load prediction model at different time scales, the weights of static attribute parameters, dynamic operating parameters, external environment parameters, or user behavior parameters in the input historical multidimensional feature data are dynamically adjusted based on the current training time scale.

5. The method for coordinated control of flexible loads in building complexes based on artificial intelligence according to claim 4, characterized in that, The multiple timescales include at least: First time scale, second time scale, and third time scale; The duration of the first time scale is shorter than that of the second time scale, and the duration of the second time scale is shorter than that of the third time scale; The weights of static attribute parameters, dynamic runtime parameters, external environment parameters, or user behavior parameters in the input historical multidimensional feature data are dynamically adjusted based on the current training timescale, including: If the current training timescale is the first timescale, increase the weights of dynamic running parameters and user behavior parameters in the input historical multidimensional feature data; If the current training timescale is the second timescale, increase the weights of the dynamic running parameters and external environment parameters in the input historical multidimensional feature data, and the weights of the dynamic running parameters and external environment parameters are the same. If the current training timescale is the third timescale, increase the weight of the external environment parameters in the input historical multidimensional feature data.

6. The method for coordinated control of flexible loads in building complexes based on artificial intelligence according to claim 5, characterized in that, Output load forecast data for at least one load unit of the target building over multiple time scales in the future, including: If the deviation between the load forecast data of at least one load unit of the target building in the first time scale and its load forecast data in the corresponding time period of the second time scale is greater than a preset first deviation threshold, the load forecast data of at least one load unit of the target building in the first time scale is updated with the load forecast data in the corresponding time period of the second time scale. If the deviation between the load forecast data of at least one load unit of the target building in the second time scale and its load forecast data in the corresponding time period of the third time scale is greater than a preset second deviation threshold, the load forecast data of at least one load unit of the target building in the second time scale shall be updated with the load forecast data in the corresponding time period of the third time scale.

7. The method for coordinated control of flexible loads in building complexes based on artificial intelligence according to claim 5, characterized in that, Output load forecast data for at least one load unit of the target building over multiple time scales in the future, including: For each time scale: Obtain the load forecast data and the actual load measurement data of the previous time step; The correction factor for the current load forecast data is determined based on the error between the load forecast data and the measured load data of the previous time. The load forecast data at the current moment is corrected using the aforementioned correction coefficient.

8. The method for coordinated control of flexible loads in building complexes based on artificial intelligence according to claim 7, characterized in that, The correction factor for the current load forecast data is determined based on the error between the load forecast data and the measured load data from the previous time step, including: The correction factor is determined using the following formula: in, Let be the correction factor at time t. The correction factor at time t-1 For smoothing coefficients, The load measurement data is at time t-1. This is the load forecast data at time t-1.

9. The method for coordinated control of flexible loads in building complexes based on artificial intelligence according to claim 1, characterized in that, The objective function is: in, Let be the energy consumption function. For cost function, For comfort function, Let be the power grid response function. The weights of the energy consumption function, The weights of the cost function, The weights of the comfort function, The weights are those of the grid response function; the energy consumption function, cost function, comfort function, and grid response function respectively represent the mapping relationship between the energy consumption, cost, comfort value, and grid response value of the target building and the control parameters of each load unit. The pre-constructed objective function is optimized and solved using an improved objective optimization algorithm based on the coupling relationship of various power consumption devices in the target building, including: S1. Construct a coupling factor to represent the coupling relationship between each power consumption device. The coupling factor is used to represent the degree of influence of the operation of any power consumption device on the control parameters of other power consumption devices. S2. Initialize the population by encoding the control parameters and coupling factors of each load unit into an individual, with each individual representing a solution to the objective function. S3. For each individual in the population, calculate the fitness value of the current individual using the objective function; S4. Determine the non-dominant relationship of each individual based on its fitness value, and then rank each individual in a hierarchical manner according to its non-dominant relationship. S5. Based on the hierarchical sorting results of each individual, determine the parent individuals, perform crossover and mutation operations on the parent individuals to update the individuals in the current population. If the convergence condition is met, take the individual with the best fitness value in the current population as the optimal solution of the objective function, and use the optimal solution as the collaborative control strategy of the at least one load unit; otherwise, return to step S3.

10. The method for coordinated control of flexible loads in building complexes based on artificial intelligence according to claim 1, characterized in that, The coordinated control strategy is dynamically adjusted based on the real-time operating status of the at least one load unit, including: Real-time acquisition of dynamic operating parameters, external environmental parameters, and user behavior parameters for each load unit corresponding to the target building; If, through a pre-built conflict resolution knowledge base, it is determined that at least one of the dynamic operating parameters, external environment parameters, and user behavior parameters of the at least one load unit meets the conflict conditions, the corresponding conflict resolution rules are determined through the conflict resolution knowledge base, and the current collaborative control strategy is adjusted according to the conflict resolution rules. The conflict resolution knowledge base includes at least conflict resolution rules corresponding to different conflict conditions. The conflict conditions include at least one of the following: dynamic operating parameters of at least one load unit, external environment parameters, and user behavior parameters, which must meet preset conditions.

11. A flexible load coordination and control device for building complexes based on artificial intelligence, characterized in that, include: The data acquisition module is configured to acquire multidimensional feature data of the target building, including static attribute parameters, dynamic operating parameters, external environment parameters, and user behavior parameters of the target building. The load forecasting module is configured to take the multidimensional feature data as input and output load forecasting data for at least one load unit of the target building in the future at multiple time scales through a pre-trained load forecasting model. The coordinated control strategy module is configured to optimize and solve a pre-constructed objective function based on the load forecast data of the multi-time scale using an improved objective optimization algorithm based on the coupling relationship of each power consumption device in the target building, so as to obtain the coordinated control strategy of the at least one load unit. The coordinated control strategy includes the control parameters of the corresponding load unit. The coordinated control module is configured to execute the coordinated control strategy on the at least one load unit and dynamically adjust the coordinated control strategy according to the real-time operating status of the at least one load unit.

12. A machine-readable storage medium storing instructions thereon, characterized in that, When executed by a processor, this instruction causes the processor to be configured to perform the AI-based flexible load coordination control method for building complexes as described in any one of claims 1-10.

13. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the artificial intelligence-based flexible load collaborative control method for building clusters as described in any one of claims 1-10.

Images