City perception data event prediction method and system based on multi-agent cooperation

Abstract

This invention provides a method and system for predicting urban sensing data events based on multi-agent collaboration. First, it acquires multi-source urban sensing data, including traffic flow data, meteorological monitoring data, population density data, and environmental sensor data. Then, it performs distributed collaborative processing through a data sensing agent, a spatiotemporal modeling agent, a graph construction agent, and an event prediction agent. Next, the data sensing agent performs entity recognition and relationship extraction on the multi-source urban sensing data to obtain heterogeneous sensing entities and multiple types of sensing relationships. Then, the spatiotemporal modeling agent generates a spatiotemporal embedding representation that integrates time and spatial dimensions. Next, the graph construction agent constructs a spatiotemporal multi-agent graph. Finally, the event prediction agent predicts urban events and outputs the probability, location, and time of occurrence. This invention achieves deep fusion of urban sensing data and accurate event prediction through multi-agent collaboration.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and smart city, in particular to a city perception data event prediction method and system based on multi-agent collaboration. BACKGROUND

[0002] With the acceleration of urbanization process, city managers are facing increasingly complex urban event management challenges. The frequent occurrence of traffic congestion, public safety events, environmental abnormalities, large-scale gatherings and other urban events has put forward higher requirements for city fine governance. City perception data, as a digital representation of city operation status, comes from traffic sensors, weather monitoring stations, people flow statistics equipment, environmental monitoring terminals and other perception devices, and has characteristics such as multi-source heterogeneity, strong spatio-temporal correlation and large data scale. How to efficiently extract valuable information from massive city perception data and accurately predict events has become a key problem in city construction.

[0003] In the prior art, a kind of event graph construction method for spatio-temporal trend data is disclosed in Chinese patent with publication number CN117456056A. This method obtains spatio-temporal trend data of the target field by constructing a spatio-temporal trend event ontology model, and extracts entities and relationships based on the ontology model to construct an event graph. However, this method has the following defects: its graph construction process relies on the processing flow of a single data source, and does not introduce a multi-agent collaboration mechanism. When faced with large-scale multi-source heterogeneous data at the city level, the efficiency of entity alignment and relationship fusion is low, and the graph update has a long time delay, making it difficult to meet the needs of real-time event prediction.

[0004] In addition, a kind of city planning achievement and large event intelligent retrieval method based on knowledge graph is disclosed in Chinese patent with publication number CN117292639A. By integrating spatio-temporal big data and multi-modal city information construction platform to construct city spatio-temporal knowledge graph, the retrieval of city entities and large events is realized. However, this method focuses on post-retrieval rather than pre-prediction, and the static structure of its knowledge graph cannot effectively capture the dynamic change characteristics of city perception data, and lacks the ability to model different types of event scenarios. There are obvious deficiencies in prediction accuracy and response speed.

[0005] Therefore, there is an urgent need for a method and system that can integrate multi-source city perception data, support multi-agent collaborative processing, have spatio-temporal dynamic modeling capability, and accurately predict multiple types of city events. SUMMARY

[0006] In view of the above problems, the purpose of this invention is to provide a method and system for predicting urban sensing data events based on multi-agent collaboration, which can achieve accurate prediction of urban events by deeply fusing and dynamically modeling multi-source urban sensing data through a multi-agent collaboration mechanism.

[0007] The first aspect of this invention provides a method for predicting urban sensing data events based on multi-agent collaboration, comprising: Acquire multi-source urban sensing data, including traffic flow data, meteorological monitoring data, population density data, and environmental sensing data; The multi-source urban sensing data is processed in a distributed and collaborative manner by multiple intelligent agents, including a data sensing intelligent agent, a spatiotemporal modeling intelligent agent, a graph construction intelligent agent, and an event prediction intelligent agent. The data sensing agent is used to perform entity recognition and relationship extraction on the multi-source urban sensing data to obtain heterogeneous sensing entities and multi-type sensing relationships. The spatiotemporal modeling agent is used to perform spatiotemporal encoding on the heterogeneous sensing entities and the multi-type sensing relationships to generate a spatiotemporal embedded representation that integrates the time and space dimensions; The spatiotemporal multi-agent graph is constructed using the graph and based on the spatiotemporal embedding representation. The spatiotemporal multi-agent graph includes entity nodes, spatiotemporal relation edges, and dynamic attribute information. The event prediction agent is used to predict urban events based on the spatiotemporal multi-agent graph, and outputs the probability of occurrence, location and time of occurrence of urban events.

[0008] In this scheme, the step of using the spatiotemporal modeling agent to perform spatiotemporal encoding on the heterogeneous sensing entities and the multi-type sensing relationships to generate a spatiotemporal embedding representation that integrates time and space dimensions includes: Each entity in the heterogeneous sensing entity is separately encoded in terms of spatial location and time sequence. The multi-type perceptual relationships are aggregated using a graph attention network to obtain relationship embedding vectors. The spatial location encoding, the time series encoding, and the relation embedding vector are fused using a multi-head cross-attention mechanism to generate the spatiotemporal embedding representation.

[0009] This plan also includes: The data-sensing intelligent agent establishes a data connection with an external urban data platform and obtains supplementary data sources from the external urban data platform. The supplementary data sources include social media data, public safety data, and municipal operation data. The supplementary data source is aligned with the multi-source city sensing data to unify entity identification and semantic representation in different data sources.

[0010] In this scheme, the step of using the event prediction agent to predict urban events based on the spatiotemporal multi-agent graph includes: Obtain the dynamic attribute change trend of each entity node in the spatiotemporal multi-agent graph; When the rate of change of entity node attributes within the target area exceeds a preset abnormal threshold, the abnormal event warning mode is triggered. In the abnormal event warning mode, causal correlation analysis of multi-hop neighbor nodes is performed on the target area to identify potential event propagation paths.

[0011] This plan also includes: Data quality assessment is performed on the multi-source urban sensing data to identify data segments with missing data, noise interference, and acquisition delays; The quality assessment results of the data fragments are used as the basis for adjusting the attention weights of the data-perceiving agent, thereby reducing the impact weight of low-quality data fragments on the map construction.

[0012] This plan also includes: The collaboration strategy and resource allocation ratio among the multiple intelligent agents are dynamically adjusted according to the type of the predicted target. When the predicted target type is a traffic congestion event, the weight of the traffic flow data in the map construction is increased; When the predicted target type is a public safety event, the weight ratio of the population density data and the public safety data in the map construction is increased.

[0013] A second aspect of the present invention provides a multi-agent collaborative event prediction system for urban sensing data, comprising a memory and a processor. The memory includes a program for a multi-agent collaborative event prediction method for urban sensing data, which, when executed by the processor, performs the following steps: Acquire multi-source urban sensing data, including traffic flow data, meteorological monitoring data, population density data, and environmental sensing data; The multi-source urban sensing data is processed in a distributed and collaborative manner by multiple intelligent agents, including a data sensing intelligent agent, a spatiotemporal modeling intelligent agent, a graph construction intelligent agent, and an event prediction intelligent agent. The data sensing agent is used to perform entity recognition and relationship extraction on the multi-source urban sensing data to obtain heterogeneous sensing entities and multi-type sensing relationships. The spatiotemporal modeling agent is used to perform spatiotemporal encoding on the heterogeneous sensing entities and the multi-type sensing relationships to generate a spatiotemporal embedded representation that integrates the time and space dimensions; The spatiotemporal multi-agent graph is constructed using the graph and based on the spatiotemporal embedding representation. The spatiotemporal multi-agent graph includes entity nodes, spatiotemporal relation edges, and dynamic attribute information. The event prediction agent is used to predict urban events based on the spatiotemporal multi-agent graph, and outputs the probability of occurrence, location and time of occurrence of urban events.

[0014] In this scheme, the step of using the spatiotemporal modeling agent to perform spatiotemporal encoding on the heterogeneous perceptual entities and the multi-type perceptual relationships to generate a spatiotemporal embedding representation that integrates time and spatial dimensions specifically includes: Each entity in the heterogeneous sensing entity is separately encoded in terms of spatial location and time sequence. The multi-type perceptual relationships are aggregated using a graph attention network to obtain relationship embedding vectors. The spatial location encoding, the time series encoding, and the relation embedding vector are fused using a multi-head cross-attention mechanism to generate the spatiotemporal embedding representation.

[0015] This plan also includes: The data-sensing intelligent agent establishes a data connection with an external urban data platform and obtains supplementary data sources from the external urban data platform. The supplementary data sources include social media data, public safety data, and municipal operation data. The supplementary data source is aligned with the multi-source city sensing data to unify entity identification and semantic representation in different data sources.

[0016] A third aspect of the present invention provides a computer-readable storage medium comprising a program for a method of predicting urban sensing data events based on multi-agent collaboration, wherein when the program is executed by a processor, it implements the steps of the method of predicting urban sensing data events based on multi-agent collaboration as described in any of the preceding claims.

[0017] This invention discloses a method and system for predicting urban sensing data events based on multi-agent collaboration. Through the distributed collaborative work of data sensing agents, spatiotemporal modeling agents, graph construction agents, and event prediction agents, it achieves deep fusion and spatiotemporal dynamic modeling of multi-source urban sensing data, constructs a spatiotemporal multi-agent graph containing entity nodes, spatiotemporal relationship edges, and dynamic attribute information, and on this basis, enables accurate prediction of the probability, location, and time of urban events, providing intelligent decision support for refined urban governance and emergency response. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope.

[0019] Figure 1 A flowchart of a method for predicting urban sensing data events based on multi-agent collaboration according to the present invention is shown. Figure 2 The flowchart of a spatiotemporal coding and feature fusion method provided by an embodiment of the present invention is shown; Figure 3 This invention provides a flowchart of a method for establishing a data source by acquiring external city data from an embodiment of the present invention. Figure 4 The diagram shows a block diagram of a city-sensing data event prediction system based on multi-agent collaboration according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Unless otherwise defined, all terms (including technical and scientific terms) used in embodiments of this invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in a common dictionary shall be interpreted as having a meaning consistent with their meaning in the context of the relevant art, and not as being interpreted in an idealized or highly formalized sense, unless expressly defined in this embodiment of the invention.

[0022] The terms "first," "second," and similar words used in the embodiments of this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "an," "a," or "the" do not indicate a quantity limitation, but rather indicate the presence of at least one. Similarly, terms such as "including" or "comprising" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. The steps preceding or following the steps in the method of the embodiments of this invention are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0023] In addition, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0024] Figure 1 The flowchart of a method for predicting urban sensing data events based on multi-agent collaboration according to the present invention is shown.

[0025] like Figure 1 As shown, the first aspect of this invention discloses a method for predicting urban sensing data events based on multi-agent collaboration, the method comprising: S102, acquire multi-source urban sensing data, including traffic flow data, meteorological monitoring data, population density data and environmental sensing data; S104, the multi-source urban sensing data is processed in a distributed and collaborative manner by multiple intelligent agents, including a data sensing intelligent agent, a spatiotemporal modeling intelligent agent, a graph construction intelligent agent, and an event prediction intelligent agent; S106, The data sensing agent is used to perform entity recognition and relationship extraction on the multi-source urban sensing data to obtain heterogeneous sensing entities and multi-type sensing relationships. S108, The spatiotemporal modeling agent is used to perform spatiotemporal encoding on the heterogeneous sensing entities and the multi-type sensing relationships to generate a spatiotemporal embedding representation that integrates the time dimension and the spatial dimension; S110, using the graph to construct an intelligent agent, a spatiotemporal multi-agent graph is constructed based on the spatiotemporal embedding representation. The spatiotemporal multi-agent graph includes entity nodes, spatiotemporal relationship edges, and dynamic attribute information. S112, the event prediction agent is used to predict urban events based on the spatiotemporal multi-agent map, and the probability of occurrence, location and time of occurrence of urban events are output.

[0026] It should be noted that in this embodiment, the urban sensing data comes from various sensor devices and data acquisition terminals deployed in the city. For example, loop detectors and cameras at intersections provide traffic flow data, meteorological stations provide meteorological monitoring data such as temperature, humidity, and precipitation, mobile phone signaling data and wireless fidelity probes provide crowd density data, and environmental monitoring stations provide environmental sensing data such as air quality and noise levels. These data vary greatly in terms of collection frequency, spatial granularity, and semantic level. Traffic data may be updated on a second-by-second basis, meteorological data is usually updated on a minute-by-minute or hour-by-hour basis, and the update interval for crowd density data is also different. In actual deployment, the data sensing agent is responsible for cleaning these heterogeneous data first, unifying the timestamps to the same reference clock, and unifying the spatial coordinates to the same coordinate system. In the entity recognition phase, the data-sensing agent extracts urban entities from various data sources, including road segments, intersections, monitoring stations, building areas, and bus stops. Relationship extraction identifies the connections between these entities, such as topological relationships like roads connecting to intersections or monitoring stations located within building areas, and statistical relationships like a positive correlation between traffic flow on one road segment and traffic flow on another. The key to multi-agent collaboration lies in the data-sensing agent passing the processed structured data to the spatiotemporal modeling agent via a message queue. The latter then passes it to the graph construction agent, and finally, the event prediction agent performs inference and prediction based on the complete graph. This pipeline-style collaboration is better than cramming all functions into a single model; each agent can be upgraded independently and operate in parallel, significantly improving the overall system throughput and maintainability. This approach enables fully automated processing of urban sensing data from collection to prediction.

[0027] Figure 2 A flowchart of a spatiotemporal coding and feature fusion method provided by an embodiment of the present invention is shown.

[0028] like Figure 2 As shown in the embodiment of the present invention, the step of using the spatiotemporal modeling agent to perform spatiotemporal encoding on the heterogeneous perceptual entities and the multi-type perceptual relationships to generate a spatiotemporal embedding representation that integrates the time and spatial dimensions specifically includes: S202, spatial location encoding and time sequence encoding are performed on each entity in the heterogeneous sensing entity; S204, The multi-type perceptual relations are aggregated using a graph attention network to obtain relation embedding vectors; S206, the spatial location encoding, the time series encoding, and the relation embedding vector are fused using a multi-head cross-attention mechanism to generate the spatiotemporal embedding representation.

[0029] It should be noted that in this embodiment, the spatial location encoding uses a learnable location embedding method, mapping the latitude and longitude coordinates of each city entity into a fixed-dimensional vector representation through a multilayer perceptron. However, unlike the simple node embedding in ordinary graph neural networks, this spatial encoding also considers the geographical distance decay effect between entities. The farther apart two entities are, the weaker their spatial interaction strength. This decay relationship can be modeled using a Gaussian kernel function. In actual parameter tuning, it was found that an decay coefficient between 0.01 and 0.05 had better performance, as verified by experiments. Larger or smaller values ​​would lead to loss of spatial resolution or amplification of noise. For time series encoding, a temporal convolutional network is used, encoding the attribute change sequence of each entity within the most recent time windows into a fixed-length vector. Taking traffic flow as an example, the hourly traffic flow change of an intersection over the past 12 hours is a sequence of length 12. The temporal convolutional network captures both short-term abrupt changes and long-term periodic patterns simultaneously through multi-scale convolutional kernels. The generation of relation embedding vectors depends on a graph attention network. This network assigns an attention weight to each relation edge, reflecting the importance of that relation to the current prediction task. The formula for calculating the attention weight is as follows:

[0030] In this model, the attention weights alpha indices i and j represent the attention weight of entity node i on entity node j. The query matrix and key matrix are the feature matrices obtained after linear transformation of the entity nodes, respectively. The key vector dimension is the number of columns in the key matrix, used to scale the dot product result to prevent gradient vanishing. Finally, a multi-head cross-attention mechanism is used to fuse the spatial encoding, temporal encoding, and relational embedding features together. This cross-attention means that the three feature paths perform attention operations on each other, rather than simply concatenating them and passing them through a fully connected layer. The advantage is that the features can fully interact with each other during fusion, avoiding the strong signal of one path suppressing the contribution of other paths. Through the above scheme, the spatiotemporal embedding representation can simultaneously capture the spatial proximity, temporal dependence, and relational semantics between urban entities, providing high-quality node and edge features for subsequent graph construction.

[0031] According to an embodiment of the present invention, the data sensing agent establishes a data connection with an external urban data platform and obtains supplementary data sources from the external urban data platform. The supplementary data sources include social media data, public safety data, and municipal operation data. The supplementary data sources are then aligned with the multi-source urban sensing data to unify the entity identifiers and semantic representations in different data sources.

[0032] It should be noted that in this embodiment, relying solely on sensor data for event prediction lacks sufficient coverage. For example, a large-scale event or emergency will generate extensive discussions and reposts on social media, which is very helpful for prediction. However, social media data is an external data source and requires proactive integration. The data-sensing intelligent agent establishes connections with external systems such as the city's big data platform and government data centers through standardized interface protocols, periodically pulling or subscribing to push the latest supplementary data. Cross-source entity alignment is a particularly challenging aspect of the entire process. Different data sources may use completely different names for the same entity. For example, in a transportation system, it might be called the eastern section of Chang'an Avenue; in a meteorological system, it might be labeled as a monitoring station in Dongcheng District; and on social media, users might simply say it's near Wangfujing. Here, a context-semantic entity alignment method is used. First, a contextual description vector is generated for each entity, and then the semantic similarity between entities from different sources is calculated. Entity pairs with a similarity exceeding a set threshold are determined to be the same entity. In actual deployment, the threshold is set to 0.85. After alignment, data from different sources are mounted onto the same graph node, enriching the node's attribute information. In this way, each entity in the graph has multi-dimensional perceptual data support, naturally improving the accuracy of event prediction. Experimental comparisons show that after introducing cross-source data fusion, the prediction accuracy of traffic congestion events increased from 78% to 87%, and the recall rate of public safety events increased by 12 percentage points.

[0033] Figure 3 The flowchart illustrates a method for establishing a data source by acquiring external city data from an embodiment of the present invention.

[0034] like Figure 3 As shown in the embodiment of the present invention, the method of using the event prediction agent to predict urban events based on the spatiotemporal multi-agent map specifically includes: Obtain the dynamic attribute change trend of each entity node in the spatiotemporal multi-agent graph; When the rate of change of entity node attributes within the target area exceeds a preset abnormal threshold, the abnormal event warning mode is triggered. In the abnormal event warning mode, causal correlation analysis of multi-hop neighbor nodes is performed on the target area to identify potential event propagation paths.

[0035] It should be noted that in this embodiment, the dynamic attribute change trend is the core indicator continuously monitored by the event prediction agent. For example, if the traffic flow at an intersection suddenly drops from 200 vehicles per minute to 50 vehicles per minute within ten minutes, while the congestion index at adjacent intersections rises rapidly, this combination of drastic attribute changes is often a precursor to traffic accidents. The anomaly threshold is not set uniformly, but dynamically adjusted according to historical statistical data by region and time period. Traffic fluctuations in commercial areas during the evening rush hour are inherently larger, so the threshold should be correspondingly relaxed; while any abnormal fluctuations during the early morning hours should be taken seriously. In the actual system, the anomaly threshold is taken as three standard deviations of the historical average for the same period as a benchmark, and then multiplied by an adjustment coefficient between 0.8 and 1.2 according to the region type. After triggering the warning mode, the event prediction agent does not only focus on the abnormal node, but also expands outward along the relational edges of the graph, performing multi-hop neighbor searches. Focusing on only one hop narrows the perspective and easily misses connections further down the causal chain. Generally, a limit of three hops is set; causal relationships beyond three hops are usually very weak and introduce significant noise. The key to causal relationship analysis lies in distinguishing between correlation and causation: congestion at node A and node B may occur simultaneously (because they are both during peak hours), but there is no causal relationship. However, if congestion at A consistently precedes congestion at B by approximately 15 minutes, then there is reason to believe causal propagation exists. Using this approach, the event prediction agent can not only detect ongoing anomalies but also predict the evolution direction and scope of impact of events, gaining valuable lead time for urban emergency response.

[0036] According to an embodiment of the present invention, the multi-source urban sensing data is subjected to data quality assessment to identify data segments with missing data, noise interference, and acquisition delay; the quality assessment results of the data segments are used as the basis for adjusting the attention weight of the data sensing agent to reduce the impact weight of low-quality data segments on map construction.

[0037] It should be noted that the quality issues of urban sensing data are very common in actual deployments, as illustrated in this embodiment. Sensors may malfunction or experience power outages leading to data loss, electromagnetic interference can generate noise spikes, and network transmission delays can cause discrepancies between data arrival time and actual collection time. If these problems are not addressed, they will introduce erroneous information into the map, making event predictions unreliable. Data quality assessment is conducted across three dimensions: completeness (whether there are missing values), accuracy (whether the values ​​are within a reasonable range), and timeliness (how long the data was delayed). Each dimension is assigned a quality score between zero and one, and the weighted average of the three dimensions is the overall quality score of the data segment. This quality score is mapped to attention weights; data with high quality scores receive greater weight in entity recognition and relation extraction, while data with low quality scores is weakened or even ignored. It's worth noting that the strategy for handling missing data is not simple discarding, but rather filling it with the historical average of the same area during the same period. However, the quality score of the filled data is marked as lower, and the corresponding attention weight is reduced accordingly. The advantage of this approach is that it ensures the completeness of the data input without allowing the filled values ​​to overshadow the contribution of the real data to map construction. Actual operational data shows that after adding the data quality assessment mechanism, the accuracy of entity attribute values ​​in the graph improved by about 6%, and the false alarm rate of event prediction decreased by nearly a quarter.

[0038] According to embodiments of the present invention, the cooperation strategy and resource allocation ratio among the multiple intelligent agents are dynamically adjusted according to the type of the predicted target; when the predicted target type is a traffic congestion event, the weight ratio of the traffic flow data in the graph construction is increased; when the predicted target type is a public safety event, the weight ratio of the crowd density data and the public safety data in the graph construction is increased.

[0039] It should be noted that in this embodiment, the data dependencies vary greatly depending on the type of urban event. For traffic congestion events, the core signals come from road network traffic flow and vehicle speed data, with weather and crowd density serving only as supplementary references. Conversely, for public safety events such as stampede risks or mass incidents, crowd density and public safety data become crucial inputs, while traffic data becomes secondary. Using the same data processing strategy and resource allocation scheme regardless of the event being predicted would waste computing power and slow down response times. In practical applications, after receiving a prediction request, the event prediction agent first parses the target event type and then sends a strategy adjustment instruction to the graph construction agent. This instruction is essentially a redistribution scheme of the weight coefficients of various data sources. Taking traffic congestion prediction as an example, the weight coefficient of traffic flow data is increased from the default 0.3 to 0.6, weather data remains at 0.2, crowd density decreases to 0.1, and environmental sensing decreases to 0.1. Public safety prediction uses the reverse allocation. The adjustment of weight coefficients not only affects the entity attribute aggregation method in the graph construction stage but also changes the filtering thresholds for different relation types during relation extraction. For example, in traffic forecasting, the extraction of social relationship edges is weakened (as these edges are not very helpful for traffic forecasting), while the extraction of road network topology relationships and traffic flow correlation relationships is strengthened. Through this on-demand scheduling strategy, the system effectively reduces unnecessary computational overhead while ensuring forecast accuracy.

[0040] According to an embodiment of the present invention, the step of constructing a spatiotemporal multi-agent graph based on the spatiotemporal embedding representation using the graph specifically includes: constructing an event-driven sub-graph by reverse associating corresponding sensing entities with city events as central nodes; and deriving implicit event associations through the co-occurrence relationships of sensing entities in the event-driven sub-graph to complete the missing relation edges in the spatiotemporal multi-agent graph.

[0041] It should be noted that in this embodiment, the conventional approach to graph construction is a forward chain from entities to relationships to events, starting with entities and relationships and then deriving events from them. However, often, knowing that a certain type of event has occurred makes it easier to deduce which entities and relationships to focus on. For example, if a region is known to have experienced a rainstorm and flooding event, then surrounding low-lying road sections, drainage network nodes, meteorological monitoring stations, and related traffic checkpoints are all entities that need to be focused on, and the causal relationships between these entities should be strengthened. The advantage of building a graph backwards from the event center is that it can quickly identify graph substructures that are highly related to the current event type, reducing interference from irrelevant information. In addition, the co-occurrence patterns of entities in the event-driven subgraph are also valuable signals. If road sensor anomalies and high drainage pump station loads frequently occur within the same time period, even if there is no direct relationship edge between them, an implicit causal relationship edge can be deduced through co-occurrence statistics. This completion strategy does not require manual annotation and is entirely data-driven and automated, making the graph increasingly complete during continuous operation.

[0042] According to an embodiment of the present invention, the method of using the event prediction agent to predict urban events based on the spatiotemporal multi-agent map further includes: calculating the confidence score of each predicted event; generating a first-level event warning signal when the confidence score is higher than a preset high confidence threshold; generating a second-level event attention prompt when the confidence score is between the preset high confidence threshold and a preset low confidence threshold; and suppressing the output of the prediction result when the confidence score is lower than the preset low confidence threshold.

[0043] It should be noted that, in this embodiment, the confidence score is calculated based on the node classification probability output by the graph neural network, which is then processed through a fully connected layer and an activation function. The calculation formula is as follows:

[0044] The confidence score represents the predicted probability of a certain type of event occurring, the weight vector is the learnable parameters of the fully connected layer, the node feature vector is the embedded representation of that node output by the graph neural network, and the bias term is a scalar bias parameter. The core idea of ​​tiered early warning is to avoid the "boy who cried wolf" problem. If all prediction results are pushed to managers, early warning fatigue will soon occur, and truly important signals will be ignored. The high confidence threshold is set at 0.85, and the low confidence threshold is set at 0.5. Predictions exceeding 0.85 have a high probability and require immediate response, generating a Level 1 early warning and pushing it to the management terminal; predictions between 0.5 and 0.85 have a certain probability but high uncertainty and are marked as Level 2 concerns, which can be displayed only on the monitoring screen without proactive notification; predictions below 0.5 are directly suppressed and do not enter the decision-making process. The specific threshold value is determined according to the needs of the actual scenario. In highly sensitive areas such as airports and train stations, the high confidence threshold can be lowered to 0.8, while 0.85 is sufficient for general commercial areas.

[0045] According to an embodiment of the present invention, the consistency check of the spatiotemporal multi-agent graph is performed periodically; when entity attribute contradictions or relationship conflicts are detected in the spatiotemporal multi-agent graph, a graph repair process is triggered, and the attribute values ​​of the contradictory entities are re-determined through a negotiation mechanism among the multiple agents.

[0046] It should be noted that in this embodiment, inconsistencies are inevitable during the continuous operation of the graph. For example, for the same road segment node, the data perception agent may have just updated its congestion index to 8 (the highest level), but the event prediction agent, based on historical data, believes its congestion index should only be 3. If this attribute contradiction is not handled in time, subsequent predictions will be based on incorrect premises. Consistency checks are performed every 30 minutes to check whether the attribute values ​​of all nodes meet preset logical constraints. For example, a road segment's congestion index cannot be lower than 4 and simultaneously marked as severely congested, and there cannot be orders of magnitude deviation between an area's environmental quality score and sensor readings. Upon discovering a contradiction, a repair process is triggered: the relevant agents each state their data source and confidence assessment results, and the graph construction agent acts as an arbitrator to make a comprehensive judgment. Generally, timestamped data has higher priority, but if the confidence levels of the new and old data sources differ significantly, a compromise value will be given by combining both. This multi-agent negotiation mechanism is much more reliable than simply overwriting old data with new data.

[0047] According to an embodiment of the present invention, a multi-granularity spatiotemporal sensing unit is constructed to divide the urban space into sensing grids of different scales; event prediction is performed for sensing grids of different scales respectively, and the prediction results of each scale are fused across scales to generate event prediction results containing multi-granularity spatiotemporal information.

[0048] It should be noted that in this embodiment, the impact range of urban events varies greatly. A localized traffic accident may only affect a few hundred meters of road, while a rainstorm may cover the entire urban area. If only a single fixed-granularity grid is used for prediction, the accuracy for local events will be insufficient (the grid is too large), or the field of view for large-scale events will be too narrow (the grid is too small). In the actual system, three scales of perception grids are maintained simultaneously: fine-grained (500-meter grid, covering core business districts and transportation hubs), medium-grained (2-kilometer grid, covering the main urban area), and coarse-grained (10-kilometer grid, covering the entire urban area). Each granularity of grid independently performs event prediction, and then the prediction results are fused across scales through a weighted voting method. Specifically, the results of fine-grained predictions have advantages in local accuracy, while coarse-grained predictions are more reliable in judging large-scale trends. During fusion, the consistency of predictions at each granularity is first checked. If all three scales predict that a certain area will experience congestion, the confidence level is high; if the fine-grained scale predicts congestion but the coarse-grained scale shows no anomalies, it may be a localized, sporadic event, and the confidence level of the fused data will be appropriately lowered. This multi-granularity collaborative prediction scheme enables the system to not only keenly capture local micro-changes but also accurately grasp the overall operational status of the city, thus substantially enhancing the robustness of the prediction results.

[0049] Please see Figure 4 The present invention also provides a multi-agent collaborative event prediction system 4 for urban sensing data, the system including a memory 401 and a processor 402, the memory including a multi-agent collaborative event prediction method program for urban sensing data, the multi-agent collaborative event prediction method program for urban sensing data being executed by the processor to implement the following steps: Acquire multi-source urban sensing data, including traffic flow data, meteorological monitoring data, population density data, and environmental sensing data; The multi-source urban sensing data is processed in a distributed and collaborative manner by multiple intelligent agents, including a data sensing intelligent agent, a spatiotemporal modeling intelligent agent, a graph construction intelligent agent, and an event prediction intelligent agent. The data sensing agent is used to perform entity recognition and relationship extraction on the multi-source urban sensing data to obtain heterogeneous sensing entities and multi-type sensing relationships. The spatiotemporal modeling agent is used to perform spatiotemporal encoding on the heterogeneous sensing entities and the multi-type sensing relationships to generate a spatiotemporal embedded representation that integrates the time and space dimensions; The spatiotemporal multi-agent graph is constructed using the graph and based on the spatiotemporal embedding representation. The spatiotemporal multi-agent graph includes entity nodes, spatiotemporal relation edges, and dynamic attribute information. The event prediction agent is used to predict urban events based on the spatiotemporal multi-agent graph, and outputs the probability of occurrence, location and time of occurrence of urban events.

[0050] It should be noted that, in this embodiment, after the program for the city-sensing data event prediction method based on multi-agent collaboration, stored in the system's memory, is loaded and executed by the processor, the system sequentially completes the steps of acquiring multi-source city-sensing data, multi-agent distributed collaborative processing, entity recognition and relationship extraction, spatiotemporal coding, graph construction, and event prediction, according to the process described in the aforementioned method implementation. Each agent is deployed on the processor as an independent software module in the system, and data exchange and collaborative scheduling are achieved through inter-process communication mechanisms. The memory also stores the model parameters, intermediate calculation results, and graph data required for the operation of each agent, ensuring that the system can quickly recover to its most recent working state after a power outage and restart. In actual deployment, the system can run on a cloud server cluster or on an edge computing node, flexibly configured according to the city's scale and real-time requirements.

[0051] According to an embodiment of the present invention, the step of using the spatiotemporal modeling agent to perform spatiotemporal encoding on the heterogeneous perceptual entities and the multi-type perceptual relationships to generate a spatiotemporal embedding representation that integrates the time and spatial dimensions specifically includes: Each entity in the heterogeneous sensing entity is separately encoded in terms of spatial location and time sequence. The multi-type perceptual relationships are aggregated using a graph attention network to obtain relationship embedding vectors. The spatial location encoding, the time series encoding, and the relation embedding vector are fused using a multi-head cross-attention mechanism to generate the spatiotemporal embedding representation.

[0052] It should be noted that, in this embodiment, after receiving the structured data transmitted by the data-aware agent, the spatiotemporal coding module in the system processes it according to the complete process of spatial location coding, temporal sequence coding, and multi-head cross-attention fusion described in the aforementioned method implementation. Parameters such as the Gaussian decay coefficient, the scale of the temporal convolution kernel, and the number of attention heads in spatial coding are all read from a configuration file loaded in memory, supporting online adjustment without requiring a system restart. The calculation of the relation embedding vector employs parallel processing using a graph attention network, which can simultaneously calculate attention weights for multiple relation edges on a multi-core processor, improving processing speed by three to five times compared to the serial mode.

[0053] According to an embodiment of the present invention, the data sensing agent establishes a data connection with an external urban data platform and obtains supplementary data sources from the external urban data platform. The supplementary data sources include social media data, public safety data, and municipal operation data. The supplementary data sources are then aligned with the multi-source urban sensing data to unify the entity identifiers and semantic representations in different data sources.

[0054] It should be noted that in this embodiment, the system's data-sensing agent integrates a standardized data access adapter layer, supporting communication with external city data platforms through various methods such as application programming interfaces (APIs), message queues, and database connections. The cross-source entity alignment module maintains an entity mapping table in memory, recording the correspondence between the identifiers of the same entity in different data sources. Whenever a new data source is connected, the system automatically runs a batch entity alignment process, semantically matching the entities in the new data source with the entities in the existing graph. Successfully matched entities are added to the mapping table, while those that fail to match are added to the graph as new entities. This incremental entity alignment strategy ensures that the system does not need to rebuild the entire graph when connecting to a new data source, guaranteeing both system scalability and operational efficiency.

[0055] The present invention also provides a computer-readable storage medium having a computer program stored thereon, the computer-readable storage medium including a program for a method for predicting urban sensing data events based on multi-agent collaboration, the program for predicting urban sensing data events based on multi-agent collaboration being executed by a processor to implement the steps of the method for predicting urban sensing data events based on multi-agent collaboration as described above.

[0056] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting events based on urban sensing data, characterized in that, The method includes: Acquire multi-source urban sensing data, including traffic flow data, meteorological monitoring data, population density data, and environmental sensing data; The multi-source urban sensing data is processed in a distributed and collaborative manner by multiple intelligent agents, including a data sensing intelligent agent, a spatiotemporal modeling intelligent agent, a graph construction intelligent agent, and an event prediction intelligent agent. The data sensing agent is used to perform entity recognition and relationship extraction on the multi-source urban sensing data to obtain heterogeneous sensing entities and multi-type sensing relationships. The spatiotemporal modeling agent is used to perform spatiotemporal encoding on the heterogeneous sensing entities and the multi-type sensing relationships to generate a spatiotemporal embedded representation that integrates the time and space dimensions; The spatiotemporal multi-agent graph is constructed using the graph and based on the spatiotemporal embedding representation. The spatiotemporal multi-agent graph includes entity nodes, spatiotemporal relation edges, and dynamic attribute information. The event prediction agent is used to predict urban events based on the spatiotemporal multi-agent graph, and outputs the probability of occurrence, location and time of occurrence of urban events.

2. The method for predicting urban sensing data events based on multi-agent collaboration as described in claim 1, characterized in that, The step of using the spatiotemporal modeling agent to perform spatiotemporal encoding on the heterogeneous perceptual entities and the multi-type perceptual relationships to generate a spatiotemporal embedding representation that integrates the time and spatial dimensions includes: Each entity in the heterogeneous sensing entity is separately encoded in terms of spatial location and time sequence. The multi-type perceptual relationships are aggregated using a graph attention network to obtain relationship embedding vectors. The spatial location encoding, the time series encoding, and the relation embedding vector are fused using a multi-head cross-attention mechanism to generate the spatiotemporal embedding representation.

3. The method for predicting urban sensing data events based on multi-agent collaboration according to claim 1, characterized in that, The method further includes: The data-sensing intelligent agent establishes a data connection with an external urban data platform and obtains supplementary data sources from the external urban data platform. The supplementary data sources include social media data, public safety data, and municipal operation data. The supplementary data source is aligned with the multi-source city sensing data to unify entity identification and semantic representation in different data sources.

4. The method for predicting urban sensing data events based on multi-agent collaboration according to claim 1, characterized in that, The method of using the event prediction agent to predict urban events based on the spatiotemporal multi-agent graph includes: Obtain the dynamic attribute change trend of each entity node in the spatiotemporal multi-agent graph; When the rate of change of entity node attributes within the target area exceeds a preset abnormal threshold, the abnormal event warning mode is triggered. In the abnormal event warning mode, causal correlation analysis of multi-hop neighbor nodes is performed on the target area to identify potential event propagation paths.

5. The method for predicting urban sensing data events based on multi-agent collaboration according to claim 1, characterized in that, The method further includes: Data quality assessment is performed on the multi-source urban sensing data to identify data segments with missing data, noise interference, and acquisition delays; The quality assessment results of the data fragments are used as the basis for adjusting the attention weights of the data-perceiving agent, thereby reducing the impact weight of low-quality data fragments on the map construction.

6. The method for predicting urban sensing data events based on multi-agent collaboration according to claim 1, characterized in that, The method further includes: The collaboration strategy and resource allocation ratio among the multiple intelligent agents are dynamically adjusted according to the type of the predicted target. When the predicted target type is a traffic congestion event, the weight of the traffic flow data in the map construction is increased; When the predicted target type is a public safety event, the weight ratio of the population density data and the public safety data in the map construction is increased.

7. A multi-agent collaborative event prediction system for urban sensing data, characterized in that, The system includes a memory and a processor. The memory includes a program for predicting events based on multi-agent collaboration of urban sensing data. When the processor executes the program for predicting events based on multi-agent collaboration of urban sensing data, it performs the following steps: Acquire multi-source urban sensing data, including traffic flow data, meteorological monitoring data, population density data, and environmental sensing data; The multi-source urban sensing data is processed in a distributed and collaborative manner by multiple intelligent agents, including a data sensing intelligent agent, a spatiotemporal modeling intelligent agent, a graph construction intelligent agent, and an event prediction intelligent agent. The data sensing agent is used to perform entity recognition and relationship extraction on the multi-source urban sensing data to obtain heterogeneous sensing entities and multi-type sensing relationships. The spatiotemporal modeling agent is used to perform spatiotemporal encoding on the heterogeneous sensing entities and the multi-type sensing relationships to generate a spatiotemporal embedded representation that integrates the time and space dimensions; The spatiotemporal multi-agent graph is constructed using the graph and based on the spatiotemporal embedding representation. The spatiotemporal multi-agent graph includes entity nodes, spatiotemporal relation edges, and dynamic attribute information. The event prediction agent is used to predict urban events based on the spatiotemporal multi-agent graph, and outputs the probability of occurrence, location and time of occurrence of urban events.

8. The multi-agent collaborative event prediction system for urban sensing data according to claim 7, characterized in that, The step of using the spatiotemporal modeling agent to perform spatiotemporal encoding on the heterogeneous perceptual entities and the multi-type perceptual relationships to generate a spatiotemporal embedding representation that integrates the time and spatial dimensions specifically includes: Each entity in the heterogeneous sensing entity is separately encoded in terms of spatial location and time sequence. The multi-type perceptual relationships are aggregated using a graph attention network to obtain relationship embedding vectors. The spatial location encoding, the time series encoding, and the relation embedding vector are fused using a multi-head cross-attention mechanism to generate the spatiotemporal embedding representation.

9. The multi-agent collaborative event prediction system for urban sensing data according to claim 7, characterized in that, Also includes: The data-sensing intelligent agent establishes a data connection with an external urban data platform and obtains supplementary data sources from the external urban data platform. The supplementary data sources include social media data, public safety data, and municipal operation data. The supplementary data source is aligned with the multi-source city sensing data to unify entity identification and semantic representation in different data sources.

10. A computer-readable storage medium having a computer program stored thereon, the computer-readable storage medium including a program for a method for predicting urban sensing data events based on multi-agent collaboration, wherein when the program for predicting urban sensing data events based on multi-agent collaboration is executed by a processor, it implements the steps of the method for predicting urban sensing data events based on multi-agent collaboration as described in any one of claims 1 to 6.

Images