A method and system for controlling station service equipment

Abstract

The application relates to a kind of method and system for controlling station service equipment, it relates to the field of rail transit station equipment technology, the method comprises: the collected multi-source perception data is input into space-time multi-task processing model, and the model output result is obtained;Multi-source perception data includes passenger flow data and weather data;Model output result includes passenger flow prediction result, multi-device operation demand prediction result and station operation scene classification result;According to the control mode corresponding to station operation scene classification result, passenger flow prediction result and multi-device operation demand prediction result, control instruction for station service equipment is generated using MPC algorithm to adjust the operating parameters of station service equipment;Wherein, space-time multi-task processing model includes: heterogeneous graph construction module, time series decomposition module, causal space-time coding module and prediction output module, so as to improve control accuracy, reduce energy consumption and improve response lag.

Claims

1. A method for controlling station service equipment, characterized in that, The method is applied to edge nodes deployed in rail transit stations; the method includes: The collected multi-source sensing data is input into a spatiotemporal multi-task processing model to obtain the model output results; the multi-source sensing data includes passenger flow data and weather data; the model output results include passenger flow prediction results, multi-equipment operation demand prediction results, and station operation scenario classification results; Based on the control mode corresponding to the station operation scenario classification result, the passenger flow prediction result, and the multi-equipment operation demand prediction result, the MPC algorithm is used to generate control instructions for the station service equipment in order to adjust the operating parameters of the station service equipment. The spatiotemporal multi-task processing model includes: a heterogeneous graph construction module, used to construct a dynamic heterogeneous graph based on pre-configured station deployment information and passenger flow direction extracted from passenger flow data, and to obtain the causal dependencies between parameters in the multi-source sensing data issued by the cloud management platform; a temporal decomposition module, used to perform temporal decomposition on the multi-source sensing data to obtain trend features, periodic features, and residual features, and to generate weather response features using the weather data, and to concatenate the four features into an enhanced temporal feature; a causal spatiotemporal coding module, used to generate a spatiotemporal representation based on the dynamic heterogeneous graph, the causal dependencies, and the enhanced temporal feature; and a prediction output module, used to output the model output result based on the spatiotemporal representation. The causal spatiotemporal coding module includes a causal heterogeneous graph attention layer. The causal heterogeneous graph attention layer generates an attention mask based on the causal dependency relationship. This means that the attention weights in the graph attention aggregation process are constrained by the causal dependency relationship matrix issued by the cloud management platform. In the calculation formula of the graph attention aggregation, the elements of the causal mask matrix are constructed as continuous values ​​jointly determined by the basic causal strength and the scene adaptive modulation. When issuing the causal dependencies, the cloud management platform also issues the base strength values ​​for each causal relationship. ,and Represents a node i With nodes j The basic causal strength value between the corresponding perception parameters, and the basic strength value is obtained by normalizing the offset of other parameters to the [0,1] interval after applying a preset change to the parameter during the causal discovery process, which reflects the inherent strength of the causal relationship. Furthermore, during inference, the edge nodes utilize the scene embedding vector generated by the scene classification branch in the prediction output module. z scene The basic causal strength is dynamically modulated; for the node i With the node j Causal mask values ​​between The calculation method is as follows: ； in, Represents the node i With nodes j The underlying causal strength values ​​between the corresponding perception parameters; express Sigmoid The function compresses the scene modulation factor to (0,1); This represents a learnable modulation weight vector, with dimensions equal to the scene embedding vector. z scene same; Indicates the Transpose of; This represents a learnable modulation bias term used to shift the modulated signal, increasing modulation flexibility.

2. The method according to claim 1, characterized in that, The time-series decomposition module is specifically used to: extract trend features from the multi-source sensing data using multi-scale moving averages to characterize the gradual changes in passenger flow and environment; extract periodic features from the collection time corresponding to the multi-source sensing data using Fourier feature mapping to characterize the periodic patterns of passenger flow; and obtain residual features to characterize sudden changes by subtracting the trend features, periodic features, and weather response features from the multi-source sensing data.

3. The method according to claim 1, characterized in that, The dynamic heterogeneous graph consists of region nodes, device nodes, and sensor nodes, and the nodes in the dynamic heterogeneous graph are connected by region adjacency edges, passenger flow edges, sensor affiliation edges, and device service edges. The causal spatiotemporal coding module includes: The temporal encoding layer is used to extract features from the enhanced temporal features along the time dimension to obtain temporal features that characterize the temporal change patterns of each node in the dynamic heterogeneous graph. The causal heterogeneous graph attention layer is used to perform graph attention aggregation on the adjacent edges of the region, the passenger flow edges, the sensor affiliation edges, and the device service edges, respectively. During graph attention aggregation, the causal dependency relationship is used as the attention mask to obtain the aggregation features corresponding to each edge, and the aggregation features corresponding to each edge are fused into spatial features. The spatiotemporal cross-attention layer is used to fuse the temporal features and the spatial features to obtain the spatiotemporal representation.

4. The method according to claim 1, characterized in that, The spatiotemporal multi-task processing model further includes an unknown scene adaptive adaptation module, used to calculate the distance between the scene embedding vector extracted during the station operation scene classification process and the scene prototype vector sent from the cloud. When the distance exceeds a preset distance threshold, the prediction output module is updated with gradients using recently cached multi-source perception data and its corresponding model output results. If the distance recalculated after the update still exceeds the preset distance threshold, the parameters in the causal spatiotemporal coding module are further updated with gradients.

5. The method according to claim 1, characterized in that, The method further includes: The control command and the currently collected multi-source sensing data are input into the digital twin module in the edge node to obtain the inferred future environment and passenger flow status; wherein, the digital twin module is constructed based on physical equations and learnable residual networks, and it is used to simulate the state changes after executing the control command with the multi-source sensing data as the initial state; The predicted future environment and passenger flow status are compared with the actual status collected at the corresponding time, and the parameters of the prediction output module are updated online based on the error between the two.

6. The method according to claim 1, characterized in that, The station service equipment includes air conditioning equipment, escalator equipment, and electronic display equipment; the control commands include the operating frequency of the air conditioning equipment, the operating speed of the escalator equipment, and the display parameters of the electronic display equipment.

7. The method according to claim 6, characterized in that, The method of generating control commands for station service equipment using the MPC algorithm includes: Using the predicted operating demand of the multi-device as the expected output, and with the goal of minimizing the comprehensive energy consumption of the station service equipment, the optimal control sequence is solved under the constraints that the operating frequency, the operating speed, and the displayed parameters are all within the allowable range, the deviation between the actual output and the expected output of the station service equipment does not exceed a preset range, and the total energy consumption of the station service equipment does not exceed a preset upper limit.

8. The method according to claim 1, characterized in that, The causal dependency relationship is obtained by the cloud management platform through the following methods: Upon receiving historical multi-source sensing data uploaded by the edge node, a causal model characterizing the dependencies between parameters in the multi-source sensing data is used to sequentially apply preset changes to each parameter in the multi-source sensing data, thereby deducing the offset of other parameters. When applying a preset change causes the offset of other parameters to exceed a preset offset threshold, it is determined that there is a causal dependency between the parameter to which the change was applied and the parameter whose offset exceeds the preset offset threshold. Based on all parameter pairs that are determined to have causal dependencies, the causal dependencies are aggregated and generated, and then distributed to the edge nodes.

9. The method according to claim 1, characterized in that, The method further includes: When a communication interruption with the cloud management platform is detected, the inference of the spatiotemporal multitasking processing model continues to be performed using the acquired causal dependency relationship as an attention mask, and the control command is generated using the MPC algorithm based on the model output result obtained from the inference. After the communication is restored, the multi-source perception data collected during the network outage and the corresponding model output result are synchronized to the cloud management platform so that the cloud management platform can use the uploaded data to perform incremental training on the spatiotemporal multitasking processing model.

10. A system for controlling station service equipment, characterized in that, This includes edge nodes deployed at rail transit stations, the edge nodes being configured to perform the method of controlling station service equipment as described in any one of claims 1 to 9.

Images