A highway service area energy consumption prediction system

Abstract

The application discloses a kind of highway service area energy consumption prediction systems, including multi-source heterogeneous data acquisition module, space-time feature extraction module, dynamic fusion module, energy consumption prediction module, prediction effect verification module and model self-optimization module.System is integrated by service area dynamic external and internal operation data, extracts multidimensional space-time characteristics, and is weighted and fused based on the fusion weight of dynamic allocation, generates space-time fusion feature set.The feature set is input to machine learning prediction model, and the total energy consumption of service area and the refined energy consumption prediction curve of each target partition in future preset period can be output simultaneously.System compares predicted and actual energy consumption data, automatically generates deviation analysis results, and accordingly realizes the self-optimization of model parameters and feature weights, thereby improving the prediction accuracy and the adaptive ability of the system, and providing effective support for realizing energy-saving operation and intelligent management of service area.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation and energy management technology, specifically to an energy consumption prediction system for highway service areas, and more particularly to an energy consumption prediction method and system capable of short-term, refined zoning prediction and possessing self-optimization capabilities. Background Technology

[0002] With the increasing density of highway networks and the continuous growth of traffic flow, highway service areas, as key supporting facilities, are experiencing a continuous expansion in their total energy consumption, making energy conservation and emission reduction increasingly urgent. Achieving green and efficient operation of service areas relies on accurate prediction of their future energy consumption trends, thereby providing data support for optimizing energy use strategies and participating in grid demand response. Therefore, developing a high-precision energy consumption prediction system is of great significance for reducing service area operating costs and promoting energy conservation and emission reduction in the transportation sector.

[0003] Currently, there has been some research on building energy consumption prediction, and some technologies have been attempted to be applied to service area scenarios. Existing technical solutions mostly focus on using historical energy consumption data to build time series prediction models (such as ARIMA models), or using a single machine learning algorithm (such as support vector machines or simple neural networks) to make a rough prediction of overall energy consumption. These methods can reflect the basic laws of energy consumption changes over time to some extent, but their data dimensions and model considerations are relatively simple.

[0004] However, in the process of realizing this invention, the inventors discovered that these existing technologies have significant limitations. First, the predictive models rely on a single data dimension, often focusing only on historical energy consumption data itself, failing to fully consider the combined influence of multi-source heterogeneous data, such as external environment (e.g., holidays, weather, real-time traffic flow) and internal dynamics of the service area (e.g., equipment start-up and shutdown in various functional areas, pedestrian flow), resulting in insufficient prediction accuracy. Second, energy consumption in service areas varies significantly in space (e.g., the energy consumption patterns of gas stations, restaurants, and charging piles are drastically different), while existing methods typically only predict the total amount, lacking the ability to predict the refined energy consumption of each functional area, thus failing to meet the needs of precise management. Furthermore, most systems are static models, lacking effective closed-loop feedback and self-optimization mechanisms, making it difficult to adapt to dynamic changes in factors such as traffic flow and seasons in service areas, causing the model's predictive performance to degrade over time. Therefore, there is an urgent need in this field for an intelligent energy consumption prediction system that can integrate multi-source spatiotemporal data, achieve accurate regional prediction, and possess self-learning capabilities. Summary of the Invention

[0005] In view of this, the present invention provides an energy consumption prediction system for highway service areas to solve at least one of the above-mentioned technical problems.

[0006] To achieve the above objectives, in a first aspect, a highway service area energy consumption prediction system is provided, comprising: The multi-source heterogeneous data acquisition module is used to acquire dynamic external data and internal operational data of highway service areas in real time from multiple categories of data sources; The spatiotemporal feature extraction module is used to process the dynamic external data and the internal operating data, and extract multiple spatiotemporal features related to energy consumption from the processed data. The dynamic fusion module is used to perform weighted fusion of multiple spatiotemporal features based on dynamically allocated fusion weights to generate a spatiotemporal fusion feature set; The energy consumption prediction module is used to input the spatiotemporal fusion feature set into the machine learning prediction model to obtain the total energy consumption prediction curve of the service area and the energy consumption prediction curve of each target partition within a future preset time period. The prediction effect verification module is used to compare the actual collected total energy consumption data of the service area and the energy consumption data of each target partition with the total energy consumption prediction curve and the energy consumption prediction curve of each target partition, and generate prediction deviation analysis results. The model self-optimization module is used to fine-tune the parameters or update the feature weights of the machine learning prediction model based on the prediction deviation analysis results.

[0007] Secondly, a method for predicting energy consumption in highway service areas is provided, which includes the following steps: Dynamic external and internal operational data of highway service areas are acquired in real time from multiple categories of data sources; The dynamic external data and the internal operating data are processed, and multiple spatiotemporal features related to energy consumption are extracted from the processed data. Based on dynamically assigned fusion weights, multiple spatiotemporal features are weighted and fused to generate a spatiotemporal fusion feature set; The spatiotemporal fusion feature set is input into the machine learning prediction model to obtain the total energy consumption prediction curve of the service area and the energy consumption prediction curve of each target partition within a future preset time period. The total energy consumption data of the service area and the energy consumption data of each target zone that are actually collected are compared with the total energy consumption prediction curve and the energy consumption prediction curve of each target zone to generate prediction deviation analysis results. Based on the prediction bias analysis results, the parameters of the machine learning prediction model are tuned or the feature weights are updated.

[0008] Thirdly, an electronic device is provided, comprising: One or more processors; Storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement a highway service area energy consumption prediction method as described in the second aspect.

[0009] Fourthly, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements a highway service area energy consumption prediction method as described in the second aspect.

[0010] Fifthly, a computer program product is provided, comprising a computer-readable storage medium storing a computer program, wherein when the computer program is run, it executes the highway service area energy consumption prediction method described in the second aspect.

[0011] The above technical solution has the following beneficial technical effects: This invention improves the accuracy and adaptability of energy consumption prediction based on a dynamic weighted fusion mechanism of multi-dimensional spatiotemporal features, overcoming the shortcomings of traditional methods such as single data utilization and rigid models. This invention can simultaneously output refined energy consumption prediction curves for the entire service area and each functional zone, providing a direct basis for precise and differentiated energy-saving management. Through a closed-loop feedback of prediction, verification, and optimization, this invention enables the system to have a self-learning ability that evolves continuously, effectively responding to the dynamic changes in energy consumption patterns in service areas, and ultimately providing reliable technical support for smart energy management and energy conservation and consumption reduction in highway service areas. Attached Figure Description

[0012] The accompanying drawings are provided to better understand the invention and are not intended to unduly limit the scope of the invention. Wherein: Figure 1 This is a schematic diagram of the logical structure of a highway service area energy consumption prediction system according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the logical structure of the spatiotemporal feature extraction module according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the working principle of the dynamic fusion module in an embodiment of the present invention; Figure 4 This is a schematic diagram of the logical structure of the energy consumption prediction module according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the logical structure of the prediction effect verification module in an embodiment of the present invention; Figure 6 This is a schematic diagram of the logical structure of the model self-optimization module in an embodiment of the present invention; Figure 7 This is a flowchart of a method for predicting energy consumption in highway service areas according to an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of a computer system according to an embodiment of the present invention. Detailed Implementation

[0013] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0014] Example 1 like Figure 1 As shown, this embodiment of the invention provides a highway service area energy consumption prediction system. This system can be physically implemented by one or more processors deployed on a server executing stored computer program instructions, and includes: The multi-source heterogeneous data acquisition module 10 is used to acquire dynamic external data and internal operation data of the highway service area in real time from multiple types of data sources; Specifically, the multi-source heterogeneous data acquisition module 10 is used to establish communication connections with various data sources inside and outside the service area and to collect data in real time or near real time. Dynamic external data includes, but is not limited to: real-time traffic flow and vehicle type data passing through the service area obtained from the traffic management department's information system (e.g., the number of vehicles passing through the upstream ETC gantry every 5 minutes); 24-hour weather forecast data obtained from the meteorological department's API interface, including temperature, humidity, wind speed, and solar radiation intensity; and date types obtained from calendar information (e.g., weekdays, weekends, and public holidays). The internal operating data (including the associated raw data corresponding to each target zone) specifically includes: real-time energy consumption metering data of each target zone (gas station, charging station, main building, etc.) collected from the sub-meters of electricity, water, and gas meters; the equipment operating status and set parameters of the corresponding target zones obtained from the main energy-consuming equipment such as central air conditioning and lighting systems deployed in the target zones such as gas stations, charging stations, and main buildings; real-time passenger flow data of the corresponding zones collected by the people counting sensors deployed in the target zones such as the main building (including restaurants and shopping malls); and the spatial energy consumption relationship characteristics of the inter-zone energy consumption change correlation coefficient, peak energy consumption time difference, energy consumption ratio / value, and shared resource allocation ratio calculated by the collected inter-zone energy consumption metering data, passenger flow, and equipment operating data, so as to quantitatively characterize the inter-zone energy consumption linkage, constraint, or collaborative relationship. This multi-source heterogeneous data acquisition module performs preliminary cleaning on the collected raw data (including data associated with each target partition), such as handling missing values ​​and removing obvious outliers, and converts it into a unified time-series data format, providing a directly related raw data foundation for subsequent feature extraction of target partitions.

[0015] The spatiotemporal feature extraction module 20 is used to process the dynamic external data and the internal operating data, and extract multiple spatiotemporal features related to energy consumption from the processed data. Specifically, the spatiotemporal feature extraction module 20 performs in-depth processing on the pre-processed data to extract key time-dimensional features and spatial-dimensional energy consumption features that directly affect energy consumption. In the time dimension, the module extracts time-period features such as hours, peak dining times, and holiday markers from the time-series data. In the spatial dimension, based on the spatial affiliation of the data, the module classifies and integrates features directly related to energy consumption, such as equipment operating loads (e.g., gas station fuel nozzle operating power, charging station charging pile output load, main building restaurant cooking equipment load), regional environmental parameters (e.g., temperature and humidity, lighting brightness in each zone), and functionally related data (e.g., gas station traffic flow, charging station charging demand frequency, main building passenger flow density), according to the corresponding target zone (e.g., the entire service area, gas stations, charging stations, main building (including restaurants and shopping malls)). This forms a spatial-dimensional energy consumption feature subset specific to each target zone. This spatial-dimensional energy consumption feature subset includes a set of features related to energy consumption changes specific to the corresponding target zone, as well as spatial energy consumption relationship features characterizing the energy consumption interaction between target zones. In this embodiment, time-series data refers to operational and environmental data collected by the multi-source heterogeneous data acquisition module from inside and outside the service area, with clear timestamps and arranged continuously in chronological order. It is used to express the dynamic changes of relevant indicators over time. It includes both dynamic external data and the serialized results of internal operational data on a unified time axis. For example, the traffic management department's information system outputs a sequence of vehicles passing through the upstream ETC gantry of the service area every 5 minutes, along with hourly temperature, humidity, wind speed, and solar intensity forecasts for the next 24 hours provided by the meteorological department's API, and a sequence of weekday / weekend / public holiday markers corresponding to calendar information. It also includes the real-time energy consumption metering sequence of electricity, water, and gas meters in each target zone of the service area, the operating status and set parameter sequence of the main energy-consuming equipment such as central air conditioning and lighting in each target zone, and the real-time passenger flow sequence counted by the pedestrian counting sensors in areas such as the main building restaurant and shopping mall. After preliminary cleaning, this data is converted into time-series data in a unified format. Furthermore, this spatiotemporal feature extraction module can construct more complex derived features. For example, it can combine real-time traffic flow with the current time to form nighttime truck traffic flow, and combine temperature and humidity from weather forecasts to form the perceived temperature index. These features together constitute a multi-level, multi-dimensional set of spatiotemporal features.

[0016] The dynamic fusion module 30 is used to perform weighted fusion of multiple spatiotemporal features based on dynamically allocated fusion weights to generate a spatiotemporal fusion feature set; Specifically, the function of the dynamic fusion module 30 is to dynamically allocate fusion weights based on the degree of influence of different spatiotemporal features on current and future energy consumption. In a preferred embodiment, the dynamic fusion module uses an attention mechanism to achieve dynamic weight allocation. The dynamic fusion module 30 analyzes the correlation between different features and final energy consumption in historical data and combines it with real-time scenarios (e.g., in hot summer weather, the weights of air temperature and air conditioning set temperature will be significantly increased; during holidays, the weights of traffic flow and pedestrian flow will become dominant) for calculation. A weight coefficient is calculated for each feature, and then all features are weighted and summed to finally generate a spatiotemporal fusion feature set that can most effectively represent the current state and is oriented towards energy consumption prediction. This dynamic weighting method ensures that the model can adapt to changes in different seasons and date types, and has higher accuracy and robustness than the fixed-weight fusion method.

[0017] The energy consumption prediction module 40 is used to input the spatiotemporal fusion feature set into the machine learning prediction model to obtain the total energy consumption prediction curve of the service area and the energy consumption prediction curve of each target partition within a future preset time period. Specifically, the energy consumption prediction module 40 takes the aforementioned spatiotemporal fusion feature set as input and imports it into a pre-trained machine learning prediction model. This prediction model can be a deep learning model integrating a Long Short-Term Memory (LSTM) network and a fully connected layer. The LSTM unit is used to learn the energy consumption evolution law from the time-series data of the spatiotemporal fusion feature set, which includes a weighted fusion result of time period features and spatial dimension features of each target partition. The fully connected layer is responsible for outputting the predicted value. After being trained on a large amount of historical data, this prediction model can simultaneously output the predicted curve of the total energy consumption of the service area in the next 24 hours (time-series data with 15-minute intervals), as well as the energy consumption prediction curves of each target partition, such as gas stations, charging stations, and the main building, within the same time period.

[0018] The prediction effect verification module 50 is used to compare the actual collected total energy consumption data of the service area and the energy consumption data of each target partition with the total energy consumption prediction curve and the energy consumption prediction curve of each target partition, and generate prediction deviation analysis results. Specifically, the prediction performance verification module 50 establishes a closed-loop feedback starting point. After the prediction period ends, the module collects the total energy consumption data of the service area and the energy consumption data of each partition during that period. Then, it compares the actual data curve with the corresponding prediction curve point by point, calculates indicators such as the root mean square error (RMSE) and mean absolute percentage error (MAPE), and generates a quantitative prediction deviation analysis result. This prediction deviation analysis result not only includes the overall degree of deviation but also identifies specific time periods or partitions with larger deviations.

[0019] The model self-optimization module 60 is used to fine-tune the parameters or update the feature weights of the machine learning prediction model based on the prediction deviation analysis results.

[0020] Specifically, the model self-optimization module 60 initiates an optimization process based on the prediction deviation analysis results. If the deviation exceeds a preset threshold, the model self-optimization module 60 triggers a retraining procedure. It uses the latest actual energy consumption data and the corresponding spatiotemporal fusion feature set as new training samples to incrementally update the parameters of the machine learning model (i.e., parameter tuning), enabling the machine learning prediction model to track the latest changes in the energy consumption patterns of the service area. In addition, the model self-optimization module 60 also analyzes the changes in the importance of each feature in the latest data and adjusts the feature weight allocation strategy in the dynamic fusion module accordingly (i.e., feature weight update). Through this continuous prediction, verification, and optimization closed loop, the system possesses the ability to learn and evolve on its own, maintain prediction accuracy over a long period, and effectively cope with the impact of changes in traffic flow patterns, equipment upgrades, and business model adjustments in the service area.

[0021] In some embodiments, the dynamic external data includes predicted traffic flow data for a future preset time period from the transportation department and weather forecast data from the meteorological department; the internal operational data includes real-time energy consumption equipment data for each functional zone of the service area. The prediction effect verification and feedback module and the model self-optimization module constitute a closed-loop optimization circuit.

[0022] In an alternative embodiment, the dynamic external data further includes vehicle brand composition data expected to enter the service area within a future preset time period based on real-time route planning, provided by the navigation platform, and real-time accident information or construction control information affecting traffic flow, released by the traffic incident detection system. The internal operational data is further expanded to include anonymized population heat map data collected through public Wi-Fi probes in the service area, real-time transaction data of major catering facilities, and indoor environmental quality data (including carbon dioxide concentration and PM2.5 values) of specific functional areas (e.g., charging stations, large supermarkets) recorded by the building automation system. These expanded data types can more precisely reflect the actual load of energy-consuming units and the intensity of population activity, providing the prediction model with behavioral characteristic dimensions beyond traditional physical parameters, thereby improving prediction accuracy and scenario adaptability.

[0023] like Figure 2 As shown, the spatiotemporal feature extraction module 20 may specifically include: Data preprocessing unit 210 is used to preprocess the dynamic external data and the internal operating data; The time feature extraction unit 220 is used to extract multi-time-scale periodic features related to energy consumption fluctuations in the service area from the preprocessed data. The spatial energy consumption feature extraction unit 230 is used to extract spatial dimension energy consumption features from the preprocessed data. The spatial dimension energy consumption features include: spatial energy consumption relationship features that characterize the energy consumption interaction relationship between each target partition in the service area and / or the energy consumption features of each target partition itself. The time feature and the spatial dimension energy consumption feature together constitute the multiple spatiotemporal features related to energy consumption.

[0024] The spatial energy consumption relationship characteristics refer to the quantitative characteristics that characterize the energy consumption linkage, constraint, or synergy between various target zones within the service area, such as gas stations, charging stations, the main building (including restaurants and shopping malls), and public restrooms, based on physical proximity and functional interaction. This reflects that energy consumption changes in different zones are not isolated but rather exhibit interconnected patterns. For example, the spatial proximity of gas stations and charging stations leads to interaction in air conditioning heat dissipation; the intensity of this linkage is quantified by the Pearson correlation coefficient of energy consumption changes in the two zones. Peak energy consumption during peak dining hours in the main building's dining area causes a delayed increase in hot water supply energy consumption in public restrooms; this temporal synergy is characterized by the time difference between the peak energy consumption in the two zones. Energy storage devices, as shared energy devices, need to distribute their discharge between charging stations and the main building; this energy consumption constraint is quantified by the allocation ratio of shared resources between the two zones. During peak load periods in the overall service area, the energy consumption ratios of gas stations and the main building exhibit an inverse relationship; this load regulation relationship is reflected by the dynamic changes in the energy consumption ratio between the two zones. The spatial energy relationship characteristics specifically include the energy consumption ratio of a region, which represents the contribution of a single region to the overall energy consumption; the energy consumption ratio, which reflects the relative correlation of energy consumption scale between regions; the correlation coefficient of energy consumption change, which quantifies the intensity of time-series linkage; the energy consumption peak time difference, which represents the time-series coordination relationship; the proportion of shared resource allocation, which reflects the energy consumption constraint relationship; and the change of dynamic energy consumption ratio between regions, which reflects the load regulation relationship.

[0025] Specifically, the preprocessing performed by the data preprocessing unit 210 includes: data cleaning, missing value imputation, and alignment and standardization of data from different data sources along the time dimension. The time feature extraction unit 220 extracts multi-timescale periodic features, including: basic periodic features, composite time features, and derived trend features; the basic periodic features include the hour of the day, the day of the week, the day of the year, and whether it is a statutory holiday; the composite time features include classification features based on hourly information, such as peak dining hours or nighttime periods; the derived trend features include calculating the rate of change of the current data relative to historical data from the same period. These features collectively enable the model to learn the differences in typical energy consumption patterns between weekdays and weekends, daytime and nighttime, and weekdays and holidays. The spatial energy consumption relationship features extracted by the spatial energy consumption feature extraction unit 230 can be quantified in the following ways: calculating the proportion of energy consumption of each target partition in the total energy consumption of the service area, representing the correlation contribution of a single target partition to the overall energy consumption of the service area; calculating the energy consumption ratio between any two target partitions, reflecting the relative correlation of energy consumption scale between target partitions; and / or calculating the correlation coefficient of energy consumption change and the time difference of energy consumption peak occurrence between any two target partitions, quantifying the temporal linkage strength and dynamic response relationship of energy consumption between sub-partitions.

[0026] The spatiotemporal feature extraction module 20 can comprehensively characterize the spatiotemporal evolution of energy consumption in the service area by collaboratively extracting the temporal periodic features and spatial correlation features of energy consumption data. This provides richer and more accurate feature inputs for subsequent fusion and prediction, ultimately improving the accuracy and generalization ability of the energy consumption prediction model.

[0027] like Figure 3 As shown, the dynamic fusion module 30 is specifically used to dynamically allocate fusion weights to each type of data source based on the data quality reliability of each type of data source and the correlation between each type of data source and historical energy consumption data; based on the fusion weights, it performs weighted calculations on the spatiotemporal features associated with the corresponding data sources to obtain multiple weighted spatiotemporal features, and merges all the weighted spatiotemporal features to generate a spatiotemporal fusion feature set. The dynamic fusion module 30 includes a data quality assessment unit 310 for calculating data quality reliability, a correlation analysis unit 320 for calculating the correlation of historical energy consumption data, a dynamic weight allocation unit 330 for allocating fusion weights to each type of data source, a weighted calculation unit 340 for generating weighted spatiotemporal features, and a feature fusion unit 350 for outputting the spatiotemporal fusion feature set.

[0028] The evaluation indicators for data quality reliability may include data integrity, stability, and timeliness. The correlation with historical energy consumption data is determined by calculating the Pearson correlation coefficient between spatiotemporal features and historical energy consumption values, or by calculating feature importance based on a machine learning model. The dynamic allocation of fusion weights may specifically include: calculating initial weights based on the weighted product of data quality reliability scores and historical correlation scores, and normalizing all initial weights. The fusion of all weighted spatiotemporal features is achieved through concatenation or weighted averaging. Specifically, it involves concatenating the weighted spatiotemporal features corresponding to each data source according to feature dimensions to form a unified feature vector, or performing a weighted average according to their fusion weights to obtain a comprehensive feature representation, thereby generating a spatiotemporal fusion feature set for subsequent prediction. This dynamic fusion mechanism, through dynamic weighting of data quality and prediction contribution in a dual-dimensional manner, enables the model to adaptively increase the decision weights of high-reliability, high-correlation features. This ensures that the feature set input to the prediction model remains optimal even when data source quality fluctuates or the external environment changes abruptly, ultimately enhancing the system's prediction robustness and accuracy in real-world complex scenarios.

[0029] In some embodiments, the dynamic fusion module first dynamically allocates fusion weights. This process is based on a two-dimensional quantitative evaluation. The first dimension is data quality reliability. The module evaluates the data quality of each data source category in real-time or near real-time, with evaluation metrics including, but not limited to, data integrity (whether there are consecutive missing values), stability (whether numerical fluctuations are within a reasonable range), and timeliness (whether data latency is within an allowable threshold). For example, the data quality reliability from high-precision electricity meters is generally higher than that of pedestrian density data estimated based on Wi-Fi probes. The second dimension is the correlation with historical energy consumption data. The module analyzes historical data to calculate the statistical correlation strength between each spatiotemporal feature and the final energy consumption value, such as using the Pearson correlation coefficient or feature importance calculated based on machine learning models (e.g., random forests). The higher the correlation, the greater the predictive value of the feature. After completing the evaluation, the dynamic fusion module calculates a comprehensive weight for each spatiotemporal feature. This comprehensive weight is a weighted product of the data quality reliability score and the historical correlation score. Subsequently, the initial weights of all features are normalized to ensure that the sum of all weights is 1. This weighting is dynamic. For example, during thunderstorms, the reliability of meteorological data (due to its significant impact on air conditioning load) and its instantaneous correlation with energy consumption both increase significantly, leading to a dynamic increase in their weights. During holidays, however, data such as traffic flow and transaction volume become dominant. Next, the dynamic fusion module performs weighted calculations and feature fusion. It uses the calculated dynamic weights to weight each spatiotemporal feature vector, generating corresponding weighted spatiotemporal features. Finally, all weighted feature vectors are concatenated or weighted averaged to generate a unified spatiotemporal fusion feature set that optimally represents the current state and is most effective for prediction. Through this mechanism, this module ensures that higher-quality features more relevant to current energy consumption prediction occupy a more significant position in the final fused features, thereby improving the input data quality for subsequent machine learning prediction models.

[0030] like Figure 4As shown, the energy consumption prediction module 40 specifically includes: a spatiotemporal input encoding unit 410, used to reconstruct features from the spatiotemporal fusion feature set, identify the degree of influence of different time scales and different spatial partitions on future energy consumption through an attention mechanism, and generate spatiotemporal encoding features with fusion attention weights based on the degree of influence; a partition prediction unit 420, used to input the spatiotemporal encoding features into independent sub-prediction models constructed for each target partition, which are then calculated by each sub-prediction model to output the energy consumption prediction sequence for each target partition; and a comprehensive energy consumption prediction unit 430, used to predict the total energy consumption prediction value of the service area within a preset time period based on the energy consumption prediction sequences of multiple target partitions, and output the total energy consumption prediction curve that changes over time. This technical solution, through a two-way mechanism of independent partition modeling and overall collaborative optimization, ensures the accuracy of energy consumption prediction for each functional area while achieving structural and temporal unity in the overall energy consumption prediction of the service area, effectively improving the collaborative accuracy and system reliability of energy consumption prediction in complex scenarios.

[0031] Specifically, the spatiotemporal input coding unit 410 works by using an attention mechanism combined with historical data correlation analysis to identify and quantify the impact of different time scales and spatial partitions on future energy consumption, and finally generate spatiotemporal coding features that integrate this impact. The specific implementation is as follows: First, it receives the spatiotemporal fusion feature set output by the dynamic fusion module. This feature set includes time dimension features (e.g., real-time time period, weekday / holiday identifier, seasonal features), spatial dimension features (e.g., location attributes of each target partition (relative coordinates, distance from main roads), equipment distribution density (number of charging piles / partition area, air conditioner indoor unit installation density), and building envelope features (external wall insulation coefficient, window area ratio). The system extracts multi-dimensional features, including energy consumption-related attributes (such as external temperature and humidity, traffic flow, and the operating load of equipment in different zones). Next, feature reconstruction is performed, dividing the time dimension into three time scales: short-term (15 minutes to 2 hours), medium-term (2 to 12 hours), and long-term (12 to 24 hours). The spatial dimension is divided according to the service area's function into target zones such as catering service area, passenger accommodation area, electric vehicle charging area, public restrooms, and office management area. Multi-time scale features and specific energy consumption-related features for each target zone are extracted (e.g., gas station traffic flow, fuel nozzle operating power, main building pedestrian flow, charging station charging pile output load, main building restaurant cooking equipment load, and public restroom hot water supply flow). A structured feature matrix was generated through dimensionality splitting and recombination (the dimension was set to [time step × number of partitions × feature dimension], where the time step was 5 minutes / step and the feature dimension was 256-dimensional). Subsequently, a multi-head self-attention mechanism (setting the number of attention heads to 8) was used to construct an attention computation network, mapping the feature matrix to a 512-dimensional high-dimensional space. Attention scores between features at different time scales and features in each partition were calculated using the scaled dot product attention formula, thereby generating an attention weight matrix. This attention score and attention weight matrix are quantitative representations of the degree of influence; higher scores / weights indicate stronger influence, such as the correlation strength between features and historical energy consumption, data quality, and partitioning. Inter-coupling relationships were investigated. The attention weights in the dining area during weekday morning rush hour (7:00-9:00) and the charging area during holiday afternoon (13:00-15:00) (similar to the long-term seasonal scale feature weight of 0.35 during the high-temperature period in summer, the feature weight of 0.28 during the middle of the morning rush hour, and the feature weight of 0.41 for the main building area during weekday morning rush hour) were significantly higher than those in other time periods / areas. Finally, the attention weight matrix was multiplied element-wise with the reconstructed feature matrix, and then the features were fused through a fully connected layer with the ReLU activation function to generate spatiotemporal encoded features. These features not only retained the core spatiotemporal information but also improved the accuracy of subsequent predictions by strengthening the feature representation of high-impact factors.

[0032] In specific implementation, the partition prediction unit 420 is implemented as follows: For each target partition, such as the catering service area and the guest accommodation area, a customized sub-prediction model is independently constructed. The selection of the sub-prediction model is determined based on the energy consumption characteristics of each partition. For example, the catering area uses the Transformer sub-model (6 encoder layers and 2048 feedforward neural network dimensions) because the cooking equipment starts and stops frequently and the energy consumption fluctuates greatly. The charging area uses the bidirectional LSTM sub-model (128 hidden layer dimensions, 3 layers, and a random inactivation rate of 0.2) because the load is strongly correlated with the traffic flow and the time sequence is prominent. The public restrooms and office management areas use the XGBoost sub-model (100 decision trees and a maximum tree depth of 6) because the energy consumption is stable and fluctuates less. The training process for each sub-prediction model is conducted independently: using historical 365 days of corresponding partition energy consumption data as labels, and partition-specific spatiotemporal coding features generated by the spatiotemporal input coding unit as input, the Adam optimizer (with an initial learning rate of 0.001, decaying by 10% every 10 rounds) is used for iterative training until the mean squared error (MSE) of the validation set converges to a preset threshold (≤0.01). In the prediction phase, the real-time generated spatiotemporal coding features are split by partition and input into the corresponding sub-prediction models. Each sub-model outputs the energy consumption prediction sequence for the corresponding partition through forward computation. The sequence time step is consistent with the feature reconstruction phase (5 minutes / step), and the preset prediction duration is 24 hours, meaning that each partition outputs energy consumption prediction values ​​for 288 time nodes.

[0033] In specific implementation, the integrated energy consumption prediction unit 430 follows these steps: First, it receives the energy consumption prediction sequences output by each partition prediction unit and constructs a sequence matrix (dimension: [number of partitions × prediction time step]). For each prediction time node, it uses an element-wise accumulation algorithm to calculate the sum of all partition energy consumption prediction values. For example, the total energy consumption prediction value at time node t (t∈[1,288]) is E_total(t)=ΣE_i(t) (where E_i(t) is the energy consumption prediction value of partition i at time node t, i∈[1,5], corresponding to 5 target partitions), thus generating the total energy consumption prediction sequence (dimension: [1×288]). To generate a continuous total energy consumption prediction curve, a cubic spline interpolation algorithm is used to smooth the total energy consumption prediction sequence, eliminating numerical jumps at discrete time nodes and ensuring the curve meets visualization and practical application requirements. Simultaneously, the energy consumption prediction sequences for each partition are retained, and the same interpolation method is used to generate the energy consumption prediction curves for each partition. Ultimately, the integrated energy consumption prediction unit outputs two forms of prediction results: one is structured data (containing a prediction sequence array of total energy consumption and energy consumption of each zone, supporting storage and subsequent analysis); the other is a visualization curve (with time as the horizontal axis and energy consumption value as the vertical axis, generated in SVG format, which can be directly connected to the front-end display module of the service area energy consumption management platform), realizing the synchronous prediction output of total energy consumption and zone energy consumption.

[0034] In another alternative embodiment, each independent sub-prediction model in the partition prediction unit 420 adopts a hybrid model structure composed of a temporal neural network and a graph convolutional network; the temporal neural network is used to obtain the temporal dependencies within the target partition, and the graph convolutional network is used to obtain the spatial interaction relationships between different target partitions; the sub-prediction model obtains the energy consumption prediction sequence of each target partition within a future preset time period based on the temporal dependencies and the spatial interaction relationships. The comprehensive energy consumption prediction unit is used to predict the total energy consumption prediction value of the service area within a future preset time period based on the energy consumption prediction sequences of multiple target partitions through weighted superposition or nonlinear fusion, and outputs a total energy consumption prediction curve that changes over time.

[0035] Specifically, the temporal neural network is a gated recurrent unit network with a hidden layer dimension of 128 and 3 layers, employing a dropout method with a dropout rate of 0.25. The adjacency matrix of the partitioned spatial interaction graph upon which the graph convolutional network is based is constructed using the inverse of the physical distance between each target partition and the functional correlation. The functional correlation is determined based on the inter-partition flow transfer coefficient or usage frequency correlation. The graph convolutional network has 2 hidden layers with a LeakyReLU activation function and a negative slope of 0.01. The sub-prediction model concatenates the temporal features output by the temporal neural network with the spatial interaction features output by the graph convolutional network, and maps them through a fully connected layer to obtain the energy consumption prediction sequence for each target partition. The hybrid model is trained using the AdamW optimizer with a learning rate of 0.0008 and a weight decay coefficient of 0.001. The training process aims to minimize the mean squared error between the predicted and true values ​​until the mean squared error of the validation set does not exceed 0.015. The pedestrian flow transfer coefficient is obtained by dividing the actual number of pedestrians flowing from target zone A to target zone B within a preset time period by the total number of pedestrians flowing out of target zone A within that time period. The frequency correlation is obtained by extracting the unit time usage frequency time series data of each target zone (e.g., number of charging times per 5 minutes, number of dining times) and calculating the Pearson correlation coefficient of this time series data between any two zones.

[0036] Specifically, the weights in the weighted superposition method are jointly determined by static base weights calculated based on historical energy consumption ratios and dynamically adjusted weights generated by a logistic regression model based on real-time spatiotemporal features. The dynamically adjusted weights serve as correction coefficients for the static base weights. The nonlinear fusion method is implemented using a lightweight multilayer perceptron, which includes an input layer, at least one hidden layer, and an output layer. It takes the energy consumption prediction sequences of the aforementioned target partitions as input and outputs the total energy consumption prediction value for the service area. The comprehensive energy consumption prediction unit is specifically used to automatically select between weighted superposition and nonlinear fusion methods based on the service area energy consumption fluctuation coefficient. When the energy consumption fluctuation coefficient is not greater than 0.3, the weighted superposition method is used; when the energy consumption fluctuation coefficient is greater than 0.3, the nonlinear fusion method is used. The comprehensive energy consumption prediction unit also smooths the generated total energy consumption prediction sequence and outputs a structured total energy consumption prediction data sequence and / or a visualized total energy consumption prediction curve. The smoothing process uses a cubic spline interpolation algorithm.

[0037] like Figure 5 As shown, the prediction effect verification module 50 may include: The error measurement unit 510 is used to calculate the prediction deviation index at multiple time scales based on the total energy consumption prediction curve and the energy consumption prediction curve of each target zone, and to construct and output a multi-scale error feature vector based on the numerical combination of the prediction deviation index at each time scale. The abnormal deviation identification unit 520 is used to analyze the multi-scale error feature vector using a statistical threshold model or a cluster-based anomaly detection model, identify abnormal deviation intervals, and mark each of the abnormal deviation intervals. The contribution source analysis unit 530 is used to perform sensitivity analysis on the spatiotemporal feature changes corresponding to different data sources based on the spatiotemporal fusion feature set within the abnormal deviation interval, in order to determine the influencing factors causing prediction deviations, and output prediction deviation analysis results containing the influencing factors. The prediction deviation analysis results include error magnitude, deviation source, corresponding influence weight, and influence factor of prediction deviation. This prediction effect verification module, through the quantitative calculation of prediction deviation indicators at multiple time scales, accurate identification of abnormal deviation intervals, and source analysis of deviation influence factors, can comprehensively and accurately quantify prediction errors, locate the root causes of deviations, and clarify influence weights, providing reliable data support for model self-optimization, thereby ensuring the accuracy and stability of energy consumption prediction.

[0038] Specifically, the error measurement unit 510 targets multiple time scales, including short-term, medium-term, and long-term scales. The multi-scale error feature vector is composed of different prediction deviation index values ​​at different time scales combined in a predetermined order. The influencing factors include the mean square error for characterizing the overall error, the mean absolute percentage error for characterizing the trend deviation, and the peak error factor for characterizing the local peak deviation.

[0039] Specifically, the anomaly identification unit 520 is used to automatically switch to a clustering-based anomaly detection model for anomaly identification when the anomaly identification accuracy of the statistical threshold model is lower than a predetermined threshold. The threshold of the statistical threshold model used by the anomaly identification unit is determined based on the mean and standard deviation of each dimension of the historical multi-scale error feature vector. The clustering-based anomaly detection model used by the anomaly identification unit is the DBSCAN algorithm.

[0040] Specifically, the contribution source tracing analysis unit 530 employs the Sobol global sensitivity analysis method, using the prediction deviation value within the abnormal deviation interval as the output response, and the spatiotemporal fusion feature set within the abnormal deviation interval as the input variable, to calculate the sensitivity index of each spatiotemporal feature in the spatiotemporal fusion feature set to the prediction deviation. The contribution source tracing analysis unit identifies features with a total sensitivity index greater than or equal to a predetermined threshold as key influencing factors, and normalizes their total sensitivity index as the corresponding influence weight. The source of deviation determined by the contribution source tracing analysis unit is determined by associating the key influencing factors with their corresponding data source categories.

[0041] Specifically, the peak error factor is calculated using the following formula: PEF = Σ| p - y p | / Σy p Wherein, PEF is the peak error factor; p is the set of actual energy consumption peak moments identified within the abnormal deviation interval, and the criterion for determining the peak moment is that the actual energy consumption value at that moment is not less than 80% of the average peak value of the same period in history; y p This represents the actual energy consumption at peak time p. p This represents the predicted energy consumption value at peak time p.

[0042] like Figure 6 As shown, the model self-optimization module 60 may include: Error attribution unit 610 is used to calculate the contribution of each feature dimension in the spatiotemporal fusion feature set based on the prediction deviation analysis results, using gradient sensitivity analysis method or feature contribution evaluation method based on attention mechanism, so as to identify target features or data sources that lead to increased prediction error. The adaptive parameter tuning unit 620 is used to automatically optimize the hyperparameters in the machine learning prediction model according to the error change trend corresponding to the target feature, using a Bayesian optimization algorithm, a genetic algorithm, or a parameter search strategy based on reinforcement learning, so as to obtain the model parameter combination with the minimum prediction error under the current running state. The online learning unit 630 is used to fine-tune the feature weights inside the machine learning prediction model in real time based on the latest actual energy consumption data and prediction deviation analysis results, through incremental learning or online gradient descent algorithm.

[0043] The model self-optimization module 60 achieves adaptive iterative optimization of the machine learning prediction model by accurately identifying the target features and data sources that cause prediction errors, dynamically optimizing model hyperparameters, and fine-tuning feature weights in real time. This effectively improves the model's adaptability to the dynamic operating status and environmental changes of the service area, and continuously ensures the high accuracy and stability of energy consumption prediction.

[0044] Specifically, the error attribution unit 610 is configured to execute gradient sensitivity analysis and attention-based feature contribution evaluation methods in parallel, and take the intersection of the results obtained by the two methods to determine the final target feature or data source. The gradient sensitivity analysis method used by the error attribution unit specifically includes: using the mean absolute percentage error in the prediction bias analysis results as the loss function, calculating the gradient of the loss function with respect to each feature dimension in the spatiotemporal fusion feature set through backpropagation, and using the normalized absolute value of the gradient as the feature attribution value. The attention-based feature contribution evaluation method used by the error attribution unit specifically includes: calculating the attention weight of each feature dimension in the spatiotemporal fusion feature set to the prediction error, and combining this attention weight with the influence weight in the prediction bias analysis results to generate a feature contribution score.

[0045] Specifically, the adaptive hyperparameter tuning unit 620 dynamically selects an optimization algorithm based on the error change trend corresponding to the target features. Specifically, a Bayesian optimization algorithm is used when the error shows an increasing trend, and a genetic algorithm is used when the error shows a stable or decreasing trend. In scenarios where the real-time performance requirement of the model prediction is higher than the optimization efficiency, a parameter search strategy based on reinforcement learning is adopted. The Bayesian optimization algorithm used by the adaptive hyperparameter tuning unit has a hyperparameter search space that includes at least the learning rate, hidden layer dimension, and dropout rate.

[0046] Specifically, the online learning unit 630 is used to fine-tune the feature weights by combining incremental learning and online gradient descent algorithms, and sets a weight adjustment threshold. When the amount of a single weight change exceeds the weight adjustment threshold, smoothing is performed. After fine-tuning the weights, the online learning unit determines whether to retain the updated weights through validation set performance evaluation; if the performance improvement reaches a first predetermined threshold, the weights are retained; if the performance decline reaches a second predetermined threshold, the weights are rolled back to the pre-adjustment weights. The online learning unit uses a sliding window mechanism to update the training data, and the sliding window includes recent historical data and the latest real-time data.

[0047] In a further embodiment, the system may further include: an energy consumption scheduling optimization decision module, used to generate an energy-saving scheduling strategy for the operation of equipment in the service area based on the total energy consumption prediction curve and the energy consumption prediction curve of each target partition, using linear programming, reinforcement learning or intelligent optimization algorithms.

[0048] Specifically, the energy consumption scheduling optimization decision module is used to dynamically select optimization algorithms based on the energy consumption characteristics of the service area. When the main goal is a single energy saving and the constraints are clear, a linear programming algorithm is selected; when the operating environment changes dynamically and real-time response is required, a reinforcement learning algorithm is selected; and when there are multiple optimization needs for energy saving, equipment life extension and service quality balance, an intelligent optimization algorithm is selected.

[0049] Specifically, the linear programming algorithm aims to minimize the total energy consumption cost over a predetermined future time period. Its constraints include at least one of equipment operation constraints, zoned energy consumption constraints, and service quality constraints. The reinforcement learning algorithm uses the Proximal Policy Optimization (PPO) algorithm to construct a decision model. Its state space includes predicted energy consumption, equipment operating status, and external environmental parameters. Its action space includes equipment start-up and shutdown control, power regulation, and priority allocation. Its reward function comprehensively measures energy-saving benefits, service quality losses, and equipment damage. The intelligent optimization algorithm uses the Non-Dominated Sorting Genetic Algorithm (NSGA-III) to solve for the Pareto optimal solution set with the objectives of minimizing total energy consumption, minimizing cumulative equipment operating losses, and maximizing the service quality compliance rate.

[0050] Specifically, the energy consumption scheduling optimization decision module is also used to receive actual execution data fed back by the equipment control system. When the deviation between the actual energy consumption and the predicted value exceeds a preset threshold, it triggers dynamic correction of the strategy based on rolling time-domain optimization to regenerate the energy-saving scheduling strategy. The energy-saving scheduling strategy generated by the energy consumption scheduling optimization decision module is output in the form of a structured instruction set containing equipment ID, scheduling period, operating parameters, and execution priority, and is sent to the equipment control system for automatic execution.

[0051] In a further embodiment, the system may also include a model multi-version management and rollback module, which stores multiple historical versions of the machine learning prediction model and automatically selects the historical model version with the best performance for rollback based on the model performance score when the prediction deviation analysis results show a continuous decline in model performance, so as to ensure prediction stability.

[0052] Specifically, the model multi-version management and rollback module includes: a version storage unit for structurally storing multiple historical versions of the machine learning prediction model, the stored content including at least model running files, version metadata, and performance benchmark indicators; a performance monitoring unit for calculating the comprehensive performance score of the currently running model based on the prediction deviation analysis results periodically generated by the prediction effect verification module, and triggering a rollback warning based on a predetermined performance continuous decline judgment condition; a score selection unit for performing performance quantification sorting and adaptability verification on historical versions after receiving a rollback warning, and determining the optimal adapted version; and a rollback execution unit for replacing the currently running model with the optimal adapted version using a gray-scale switching mechanism.

[0053] Specifically, the criteria for determining continuous performance decline include at least two of the following: the overall performance score decreases by more than a predetermined threshold over multiple consecutive update cycles; the latest overall performance score is lower than a predetermined percentage of the average score of all historical versions; and the core prediction deviation index exceeds a preset threshold over multiple consecutive update cycles without any convergence trend.

[0054] Specifically, the rollback execution unit employs a canary switching mechanism, which includes: gradually migrating real-time prediction requests from the currently running model to the optimal adapted version, monitoring the model response status during the migration process, and immediately terminating the switching and rolling back to the currently running model if an anomaly occurs. The model multi-version management and rollback module also includes a version maintenance unit, used to automatically clean and archive stored historical versions according to preset version retention rules, and periodically perform version integrity checks. The scoring and selection unit determines the optimal adapted version based on the comprehensive performance score of historical versions and their compatibility with the current system's feature dimensions, data sources, and model interfaces.

[0055] In a further embodiment, the system may also include an online incremental learning module, which is used to update the local parameters of the machine learning prediction model in real time based on newly collected data without retraining the complete machine learning prediction model, so as to adapt to the dynamic changes in the service area's operating status and environmental conditions.

[0056] Specifically, the incremental learning module may include: a data filtering unit for preprocessing and validating newly collected data and constructing a sliding data window; a local parameter localization unit for determining the set of local parameters in the machine learning prediction model that can be updated in real time through parameter sensitivity analysis; an incremental update unit for incrementally updating the set of local parameters based on the data in the sliding data window using an online gradient descent algorithm; and a stability verification unit for verifying the stability of the machine learning prediction model's performance after the incremental update and triggering a parameter rollback mechanism if the verification fails. Specifically, the stability verification unit verifies the stability of the machine learning prediction model's performance after the incremental update; when the verification fails, it triggers a parameter rollback mechanism to restore the local parameters of the machine learning prediction model to their state before the incremental update. Specifically, the data filtering unit performs validity verification on the newly collected data, including at least one of data quality checking, data distribution consistency testing, and information gain evaluation. The set of local parameters determined by the local parameter localization unit includes the weight parameters and feature fusion coefficients of the model output layer, but excludes the core parameters of the underlying feature extraction layer. The incremental update unit adopts a dynamic learning rate strategy, which adjusts the learning rate in real time according to changes in the loss function. The stability verification unit uses a hybrid dataset containing historical core data and newly added data for verification, and sets thresholds for multiple deviation indicators as verification standards. The online incremental learning module also includes an update monitoring unit, which records incremental update logs and monitors the performance trend of the machine learning prediction model in real time, dynamically adjusting the frequency of incremental updates based on the performance trend.

[0057] In a further embodiment, the system may further include a vehicle-road cooperative information fusion module, used to acquire real-time vehicle-road cooperative data from the vehicle-road cooperative system, the vehicle-road cooperative data including but not limited to speed limit information, lane closure information, road congestion level and traffic events; the multi-source heterogeneous data acquisition module is also used to acquire the vehicle-road cooperative data; the spatiotemporal feature extraction module is also used to extract energy consumption-related features from the vehicle-road cooperative data, for the dynamic fusion module to fuse and the energy consumption prediction module to predict.

[0058] In a further embodiment, the system may further include a multi-objective optimization module, which is used to construct a multi-objective optimization function based on prediction accuracy, energy consumption stability and cost constraints, and to guide the model self-optimization module to perform hyperparameter optimization or feature weight adjustment.

[0059] Specifically, the multi-objective optimization module may include: an objective quantization and weight configuration unit, used to quantify the three optimization objectives of prediction accuracy, energy consumption stability, and cost constraints, and dynamically configure the weights of each objective based on the real-time operation scenario of the service area; a function construction and solution unit, used to construct a multi-objective optimization function based on the quantized objectives and dynamic weights, and solve it using a multi-objective optimization algorithm to obtain a Pareto optimal solution set; a result filtering and adaptation unit, used to filter out the optimal adaptation solution that is compatible with the current model state from the Pareto optimal solution set; and a linkage execution unit, used to convert the optimal adaptation solution into a hyperparameter optimization range constraint and / or feature weight adjustment direction guide for the model self-optimization module.

[0060] Specifically, in the target quantization and weight configuration unit, the prediction accuracy target is quantified using the comprehensive deviation index P, the energy consumption stability target is quantified using the fluctuation control index S, and the cost constraint target is quantified using the economic efficiency index C. The target quantization and weight configuration unit is configured to dynamically adjust the weight distribution among the three optimization targets—prediction accuracy, energy consumption stability, and cost constraint—based on whether the service area is in a peak energy consumption period, a cost-sensitive period, or a period of severe energy consumption fluctuation. The function construction and solution unit uses the non-dominated sorting genetic algorithm (NSGA-Ⅲ) to solve the multi-objective optimization function. When the result screening and adaptation unit selects the optimal fit solution, it sequentially performs constraint filtering, weighted objective function value comparison, and model compatibility verification. The linkage execution unit transmits the hyperparameter combination in the optimal fit solution to the adaptive parameter tuning unit of the model self-optimization module to constrain its hyperparameter optimization space; and transmits the feature weight adjustment vector to the online learning unit of the model self-optimization module to guide the direction and magnitude of its feature weight adjustment. The linkage execution unit is also used to receive feedback data from the model self-optimization module and decide whether to re-trigger the multi-objective optimization process based on the optimization effect of the feedback.

[0061] The formula for calculating the comprehensive deviation index P is: P = 0.5 × MSE_norm + 0.3 × MAPE_norm + 0.2 × PEF_norm, where MSE_norm, MAPE_norm, and PEF_norm are the normalized values ​​of each deviation index, calculated by dividing the actual value of the index by the historical maximum index value, and the value range is [0,1]. The smaller P is, the higher the prediction accuracy. The formula for calculating the fluctuation control index S is: S = standard deviation of the predicted energy consumption curve / standard deviation of the historical actual energy consumption curve. S≤1 indicates that the fluctuation of the predicted curve conforms to the actual law, and the smaller S is, the stronger the stability. The economic efficiency index C is calculated as follows: C = (current parameter optimization computing cost + model operation energy consumption cost) / benchmark cost (benchmark cost is the average cost of the historical optimal parameter combination), and C ≤ 1 indicates that cost control meets the standard.

[0062] In a further embodiment, the system may further include a panoramic intelligent scheduling system module. This module works in conjunction with the energy consumption scheduling optimization decision module, and its inputs include the energy consumption prediction curve output by the energy consumption prediction module, real-time operating data acquired by the multi-source heterogeneous data acquisition module, and user service demand constraints. This module is configured to: run multiple scheduling algorithms (linear programming, heuristic search, deep reinforcement learning) in parallel, and evaluate and fuse the scheduling strategies generated by each algorithm to generate a comprehensively optimal equipment scheduling strategy, thereby achieving the best balance between minimizing total energy consumption and satisfying user needs.

[0063] Specifically, the panoramic intelligent scheduling system module, which works in conjunction with the energy consumption scheduling optimization decision module, includes: a data integration unit, used to receive and integrate energy consumption prediction curves from the energy consumption prediction module, real-time operating data from the multi-source heterogeneous data acquisition module, and user service demand constraints, to generate a unified format scheduling input dataset; a parallel algorithm scheduling unit, used to run multiple scheduling algorithms in parallel based on the scheduling input dataset, including linear programming algorithms, heuristic search algorithms, and deep reinforcement learning algorithms, to generate their respective scheduling strategies; a strategy evaluation unit, used to construct a multi-dimensional evaluation system to quantitatively score the scheduling strategies generated by the multiple scheduling algorithms; a strategy fusion unit, used to generate a comprehensive optimal scheduling strategy based on the quantitative scoring results, using a mechanism of complementary fusion of main and auxiliary strategies; and a collaborative execution unit, used to decompose the comprehensive optimal scheduling strategy into basic scheduling instructions and dynamic adjustment instructions, and to execute them collaboratively with the energy consumption scheduling optimization decision module.

[0064] Specifically, in the parallel algorithm scheduling unit, the linear programming algorithm takes minimizing total energy consumption cost as its objective function and is subject to equipment physical constraints, service demand constraints, and power supply capacity constraints; the heuristic search algorithm adopts an improved genetic algorithm, whose fitness function integrates energy consumption, service quality, and cost indicators; the deep reinforcement learning algorithm adopts a dual-delay deep deterministic policy gradient algorithm for real-time response scheduling in dynamic environments. The evaluation indicators of the policy evaluation unit include core indicators and auxiliary indicators. The core indicators include total energy consumption reduction rate, service demand satisfaction rate, and scheduling cost saving rate. The auxiliary indicators include equipment operation stability, policy executability, and emergency response redundancy. The policy fusion unit adopts a complementary fusion mechanism of main policy and auxiliary policy, specifically including: firstly, selecting the policy with the highest evaluation score as the main policy, and selecting the policy with the best performance in core indicators among the remaining policies as the auxiliary policy; and then fusing the main policy and the auxiliary policy through conflict resolution, complementary advantages, and smooth optimization to generate the comprehensive optimal scheduling policy. The collaborative execution unit divides the comprehensive optimal scheduling strategy into basic scheduling instructions and dynamic adjustment instructions: the basic scheduling instructions are executed through the energy consumption scheduling optimization decision module for routine scheduling; the dynamic adjustment instructions are dynamically issued by the panoramic intelligent scheduling system module based on real-time monitoring of triggering conditions. The panoramic intelligent scheduling system module also includes a feedback optimization unit, used to collect deviations between actual operating data and the scheduling strategy, calculate the strategy execution effect evaluation score, and dynamically adjust the scheduling strategy generation frequency or update the target weights of the multi-objective optimization module based on the evaluation score.

[0065] Among them, conflict resolution prioritizes the fulfillment of rigid constraints. For scheduling command conflicts of the same device in the main and auxiliary strategies, rigid constraint-related conflicts are directly executed according to the command that meets the constraints, while flexible constraint-related conflicts determine the optimal command through voting based on the corresponding index weights. The complementary advantages focus on the core index weakness of the main strategy, extract the optimization logic corresponding to the weak index in the auxiliary strategy, and embed it into the corresponding time node and partition of the main strategy to enhance the overall optimization effect. The smoothing optimization performs time-series smoothing processing on the fused scheduling commands by setting the maximum rate of change threshold of device operating parameters (such as power and temperature) to avoid instability of device operation caused by parameter mutations.

[0066] Those skilled in the art will clearly understand 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.

[0067] like Figure 7As shown in the figure, this embodiment of the invention also provides a method for predicting energy consumption in highway service areas, which includes the following steps: S71. Obtain dynamic external data and internal operational data of highway service areas in real time from multiple categories of data sources; S72. Process the dynamic external data and the internal operating data, and extract multiple spatiotemporal features related to energy consumption from the processed data; S73. Based on dynamically allocated fusion weights, multiple spatiotemporal features are weighted and fused to generate a spatiotemporal fusion feature set; S74. Input the spatiotemporal fusion feature set into the machine learning prediction model to obtain the total energy consumption prediction curve of the service area and the energy consumption prediction curve of each target partition within a future preset time period. S75. Compare the actual collected total energy consumption data of the service area and the energy consumption data of each target partition with the total energy consumption prediction curve and the energy consumption prediction curve of each target partition to generate prediction deviation analysis results. S76. Based on the prediction deviation analysis results, perform parameter tuning or feature weight update on the machine learning prediction model.

[0068] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-described methods for predicting energy consumption in highway service areas.

[0069] The present invention can implement all or part of the processes in the methods of the above embodiments, or it can be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0070] The present invention also provides an electronic device. The electronic device of an embodiment of the present invention includes: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the highway service area energy consumption prediction method provided by the present invention.

[0071] The following is for reference. Figure 8 This illustrates a schematic diagram of the structure of a computer system 800 suitable for implementing embodiments of the present invention in an electronic device. For example... Figure 8 As shown, the computer system 800 includes a central processing unit (CPU) 801, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 802 or programs loaded from storage section 808 into random access memory (RAM) 803. The RAM 803 also stores various programs and data required for the operation of the computer system 800. The CPU 801, ROM 802, and RAM 803 are interconnected via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0072] The following components are connected to I / O interface 805: an input section 806 including a keyboard, mouse, etc.; an output section 807 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN card, modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to I / O interface 805 as needed. A removable medium 811, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 810 as needed so that computer programs read from it can be installed into storage section 808 as needed.

[0073] In particular, according to the embodiments disclosed in this invention, the processes described in the above main step diagrams can be implemented as computer software programs. For example, embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the main step diagrams. In the above embodiments, the computer program can be downloaded and installed from a network via communication section 809, and / or installed from removable medium 811. When the computer program is executed by central processing unit 801, it performs the functions defined in the system of this invention.

[0074] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can occur depending on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A highway service area energy consumption prediction system, characterized in that, include: The multi-source heterogeneous data acquisition module is used to acquire dynamic external data and internal operational data of highway service areas in real time from multiple categories of data sources; The spatiotemporal feature extraction module is used to process the dynamic external data and the internal operating data, and extract multiple spatiotemporal features related to energy consumption from the processed data. The dynamic fusion module is used to perform weighted fusion of multiple spatiotemporal features based on dynamically allocated fusion weights to generate a spatiotemporal fusion feature set; The energy consumption prediction module is used to input the spatiotemporal fusion feature set into the machine learning prediction model to obtain the total energy consumption prediction curve of the service area and the energy consumption prediction curve of each target partition within a future preset time period. The prediction effect verification module is used to compare the actual collected total energy consumption data of the service area and the energy consumption data of each target partition with the total energy consumption prediction curve and the energy consumption prediction curve of each target partition, and generate prediction deviation analysis results. The model self-optimization module is used to fine-tune the parameters or update the feature weights of the machine learning prediction model based on the prediction deviation analysis results.

2. The system according to claim 1, characterized in that, The dynamic external data includes predicted traffic flow data and weather forecast data for a preset time period in the future; the internal operation data includes real-time energy consumption equipment data for each functional area of ​​the service area.

3. The system according to claim 1, characterized in that, The prediction effect verification and feedback module and the model self-optimization module form a closed-loop optimization circuit.

4. The system according to claim 1, characterized in that, The spatiotemporal feature extraction module specifically includes: A data preprocessing unit is used to preprocess the dynamic external data and the internal operating data; The time feature extraction unit is used to extract multi-timescale periodic features related to energy consumption fluctuations in the service area from the preprocessed data. The spatial energy consumption feature extraction unit is used to extract spatial dimension energy consumption features from preprocessed data. The spatial dimension energy consumption features include: spatial energy consumption relationship features that characterize the energy consumption interaction relationship between target partitions within the service area and / or the energy consumption features of each target partition itself.

5. The system according to claim 1, characterized in that, The dynamic fusion module is specifically used to dynamically allocate fusion weights to each type of data source based on the data quality reliability of each type of data source and the correlation between each type of data source and historical energy consumption data; based on the fusion weights, the spatiotemporal features associated with the corresponding data source are weighted and calculated to obtain multiple weighted spatiotemporal features, and all weighted spatiotemporal features are fused to generate a spatiotemporal fusion feature set.

6. The system according to claim 1, characterized in that, The energy consumption prediction module includes: The spatiotemporal input coding unit is used to reconstruct the spatiotemporal fusion feature set, identify the degree of influence of different time scales and different spatial partitions on future energy consumption through an attention mechanism, and generate spatiotemporal coding features with fusion attention weights based on the degree of influence. The partition prediction unit is used to input the spatiotemporal coding features into the independent sub-prediction models constructed for each target partition, and each sub-prediction model performs calculations and outputs the energy consumption prediction sequence for each target partition respectively. The integrated energy consumption prediction unit is used to predict the total energy consumption of the service area within a preset time period based on the energy consumption prediction sequence of multiple target partitions, and output the total energy consumption prediction curve that changes over time.

7. The system according to claim 6, characterized in that, Each independent sub-prediction model in the partition prediction unit adopts a hybrid model structure composed of a temporal neural network and a graph convolutional network; the temporal neural network is used to obtain the temporal dependency relationship within the target partition, and the graph convolutional network is used to obtain the spatial interaction relationship between different target partitions; the sub-prediction model obtains the energy consumption prediction sequence of each target partition in the future preset time period based on the temporal dependency relationship and the spatial interaction relationship.

8. The system according to claim 6 or 7, characterized in that, The comprehensive energy consumption prediction unit is used to predict the total energy consumption of the service area within a preset time period in the future by weighted superposition or nonlinear fusion based on the energy consumption prediction sequences of multiple target partitions, and output the total energy consumption prediction curve that changes over time.

9. The system according to claim 1, characterized in that, The prediction effect verification module includes: The error measurement unit is used to calculate the prediction deviation index at multiple time scales based on the total energy consumption prediction curve and the energy consumption prediction curve of each target zone, and to construct and output a multi-scale error feature vector based on the numerical combination of the prediction deviation index at each time scale. An abnormal deviation identification unit is used to analyze the multi-scale error feature vector using a statistical threshold model or a cluster-based anomaly detection model to identify abnormal deviation intervals. The contribution source analysis unit is used to perform sensitivity analysis on the spatiotemporal feature changes corresponding to different data sources based on the spatiotemporal fusion feature set within the abnormal deviation interval, so as to determine the influencing factors that cause prediction deviation, and output the prediction deviation analysis results containing the influencing factors.

10. The system according to claim 1, characterized in that, The model self-optimization module includes: The error attribution unit is used to calculate the contribution of each feature dimension in the spatiotemporal fusion feature set based on the prediction deviation analysis results, using gradient sensitivity analysis or feature contribution evaluation method based on attention mechanism, so as to identify the target feature or data source that leads to the increase of prediction error. The adaptive parameter tuning unit is used to automatically optimize the hyperparameters in the machine learning prediction model according to the error change trend corresponding to the target feature, using Bayesian optimization algorithm, genetic algorithm or parameter search strategy based on reinforcement learning, so as to obtain the model parameter combination with the minimum prediction error under the current running state. The online learning unit is used to fine-tune the feature weights within the machine learning prediction model in real time, based on the latest actual energy consumption data and prediction deviation analysis results, through incremental learning or online gradient descent algorithms.

Images