A method and system for end-to-end time series prediction of flash flood warning in small watersheds
By constructing a multi-source heterogeneous data stream with dual timestamps and an end-to-end time series prediction branch, the problem of video resource contention caused by high-frequency concurrent early warning events was solved. This enabled the effective screening of early warning events and the priority acquisition of video resources, ensuring the efficient operation of flood control emergency dispatch.
Patent Information
- Application Number
- CN202611354827.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-03
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-09-03
AI Technical Summary
In flood control emergency response scenarios where multiple departments are on joint duty, the conventional first-come-first-served dispatch mechanism leads to frequent competition for video resources by high-frequency concurrent early warning events, causing the PTZ camera to malfunction and unable to acquire high-priority early warning videos in a timely manner, delaying decision-making and resulting in people being trapped or property damage.
By constructing a multi-source heterogeneous data stream with dual timestamps, spatiotemporal interpolation alignment and missing item completion are performed. Combined with end-to-end time series prediction branches, multi-timescale rolling prediction is carried out. Spatial radius coverage query and time dimension anti-shake filtering are performed to filter effective early warning event sequences. Through hierarchical preemptive PTZ lock signaling interaction, video resources are prioritized to generate early warning handling work orders.
It enables effective screening of early warning events and priority acquisition of video resources, avoids network communication link blockage, ensures orderly queuing and accurate response of emergency dispatch channels, and ensures timely processing of high-priority early warnings.
Smart Images

Figure CN122867377B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of flash flood early warning technology, and more specifically, to a method and system for end-to-end time series prediction of flash floods in small watersheds. Background Technology
[0002] Flash floods are a common natural disaster in mountainous watersheds. To prevent flash floods and protect lives and property, flash flood early warning systems for small watersheds have been widely used. End-to-end time series prediction and video monitoring linkage are important means to improve the timeliness of early warnings. Patent application CN121963412A discloses a flash flood early warning system based on deep learning and integrated air-space-ground perception. The adopted scheme collects multi-source data through air-space-ground perception modules and integrates a physically constrained spatiotemporal attention model and edge computing to predict flash flood parameters. The effect is to build a multi-element linkage monitoring system, improving the comprehensiveness and timeliness of flash flood early warnings. Patent CN118397797B discloses a method and device for forecasting and early warning of flash floods in small watersheds, electronic equipment, and storage medium. The adopted scheme establishes a time-series water level data sample set, preprocesses and corrects abnormal water level data by combining preset thresholds, and executes early warning processing when specific fitting conditions are met. The effect is to achieve forecasting and early warning of flash floods in small watersheds and to correct abnormal water levels.
[0003] However, in flood control emergency response scenarios involving multiple departments on joint duty, when the water level at key monitoring stations fluctuates near the threshold boundary, frequent model noise disturbances can instantly generate a massive number of early warning events. If a conventional first-come, first-served dispatch mechanism is used at this time, frontline personnel often face situations where low-level warning signals (such as blue warnings) preemptively seize and lock onto the target camera pan-tilt units, preventing control of subsequent high-level warnings (such as red immediate evacuation warnings). The monitoring terminal screens are unable to retrieve video streams of the core scene for extended periods, and the on-site pan-tilt motors repeatedly start and stop under the impact of concurrent commands from multiple terminals, easily leading to mechanical jamming. Once the golden verification window after an urgent warning is missed, flood control commanders will delay issuing evacuation decisions due to a lack of visual support, ultimately resulting in irreversible consequences such as people being trapped or property damage during the rapid, minute-by-minute rise in flash floods. Summary of the Invention
[0004] In an early warning system built on a distributed intelligent agent network, discrete sensor nodes and multi-level monitoring terminals constitute a decentralized concurrent source. In the actual flood control workflow, the raw measurements generated by multi-source heterogeneous sensing devices near the critical point of water level threshold are often accompanied by high-frequency random jitter. After being amplified by a time series model prediction, the fluctuating signals are transformed into dense instantaneous early warning indicators. When a massive number of discrete early warning indicators flood into the video verification scheduling queue, each distributed intelligent agent terminal will send pan-tilt-zoom (PTZ) turning and locking commands to the target camera almost simultaneously based on the independent events received. Due to the lack of time-dimensional signal smoothing and spatial-dimensional priority isolation and control, the underlying PTG motor control board must forcibly respond to asynchronous concurrent commands from different terminals and different early warning levels. The disordered interweaving of control commands not only causes the network communication link and streaming media server to face instantaneous connection overload, but also forces the PTG physical components into a deadlock oscillation state of frequent relocation, resulting in high-priority urgent video streaming requests being forcibly suspended for a long time, blocking the closed-loop channel of early warning from algorithm deduction to manual intervention.
[0005] To overcome the shortcomings of existing technologies, such as the lack of early warning anti-shake filtering and hierarchical preemption mechanisms leading to video verification channel congestion under high-frequency concurrency, this invention provides an end-to-end time series prediction method and system for early warning of flash floods in small watersheds. By introducing time-dimensional anti-shake filtering and spatial radius coverage query on the basis of time series forecasting, it filters effective early warning event sequences and identifies candidate PTZ resources. Then, it uses hierarchical preemptive PTZ lock signaling and arbitration results from concurrent request conflict queues to control video streaming. This invention utilizes data streams with dual timestamps and a spatiotemporal alignment mechanism to ensure the accuracy of the prediction model input. At the early warning execution end, it prioritizes the verification of video monitoring resources for high-priority events, avoiding hardware mechanical losses caused by the interleaving of concurrent control commands and ensuring orderly queuing and accurate response of the emergency dispatch channel.
[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for early warning of flash floods in small watersheds based on end-to-end time series prediction includes: The raw measurement values of multi-source heterogeneous sensing devices deployed in a small watershed are collected to construct a multi-source heterogeneous raw data stream with double timestamps. The multi-source heterogeneous raw data stream with double timestamps is sequentially subjected to spatiotemporal interpolation alignment and missing term completion to generate a complete data matrix. The complete data matrix is input into the end-to-end time series prediction branch for multi-timescale rolling prediction, and the flood peak forecast result is output. Based on the flood peak forecast result, a hierarchical mapping is performed to obtain a four-level forecast and warning label. For the Level 4 forecast and early warning indicators, spatial radius coverage query and temporal anti-shake filtering are performed to obtain a candidate camera PTZ resource ranking list and a valid early warning event sequence. Based on the candidate camera PTZ resource ranking list and the valid early warning event sequence, hierarchical preemptive PTZ lock signaling interaction is performed to determine the target camera PTZ. A concurrent request conflict queue is established to output the conflict arbitration result to control the duty terminal to pull the video image from the target camera PTZ. The video image is bound with the valid early warning event sequence and the flood peak forecast result to generate an early warning handling work order.
[0007] An end-to-end time series prediction system for flash flood warning in small watersheds, used to implement the aforementioned end-to-end time series prediction method for flash flood warning in small watersheds, the system comprising: Multi-source data processing module: used to collect raw measurement values from multi-source heterogeneous sensing devices deployed in a small watershed to construct a multi-source heterogeneous raw data stream with dual timestamps, and sequentially perform spatiotemporal interpolation alignment and missing item completion on the multi-source heterogeneous raw data stream with dual timestamps to generate a complete data matrix; The flood peak forecast and early warning module is used to input the complete data matrix into the end-to-end time series prediction branch for multi-time-scale rolling prediction, output the flood peak forecast result, and perform hierarchical mapping based on the flood peak forecast result to obtain four-level forecast and early warning labels; The PTZ resource scheduling module is used to perform spatial radius coverage query and time dimension anti-shake filtering for the four-level forecast and early warning indicators, respectively to obtain the candidate camera PTZ resource sorting list and the effective early warning event sequence. Based on the candidate camera PTZ resource sorting list and the effective early warning event sequence, the module performs hierarchical preemptive PTZ lock signaling interaction to determine the target camera PTZ. Early warning and handling work order module: Used to establish a concurrent request conflict queue, output conflict arbitration results to control the duty terminal to pull video images from the target camera pan-tilt unit, and bind the video images with the valid early warning event sequence and flood peak forecast results to generate an early warning and handling work order.
[0008] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention overcomes the sampling misalignment and data gaps caused by multi-source heterogeneous raw data streams with dual timestamps and performs spatiotemporal interpolation alignment and missing term completion, providing the prediction model with strictly time-aligned and spatially complete matrix input conditions. It utilizes an end-to-end time series prediction branch to perform multi-timescale rolling predictions and hierarchical mapping to four levels of forecast and early warning indicators, transforming the complex hydrological dynamics of small watersheds into intuitive and standardized early warning response signals. By employing spatial radius coverage query and temporal dimension anti-shake filtering, it accurately identifies candidate video resources, removes model noise disturbances at threshold boundaries, and extracts... A reliable and effective sequence of early warning events is established to avoid the impact of momentary invalid alarm requests on the underlying network communication links. Based on a hierarchical preemptive PTZ lock signaling interaction and concurrent request conflict queue arbitration mechanism, control command conflicts caused by multiple terminals concurrently competing for the same target camera PTZ are resolved. A physical resource scheduling logic is established that prioritizes verification for high-alert-level events and orderly alternatives for low-alert-level events. Finally, the on-site video footage, early warning events, and flood peak forecast results are combined to form an early warning handling work order, fully opening up an end-to-end automated handling channel from front-end multi-source data perception, time-series rolling prediction, PTZ anti-conflict preemption to business verification and archiving. Attached Figure Description
[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0010] Figure 1 This is a flowchart of a small watershed flash flood early warning method for end-to-end time series prediction according to the present invention; Figure 2 This is a flowchart of the spatiotemporal alignment process for multi-source data acquisition in this invention; Figure 3 This is a schematic diagram of the asynchronous alignment of dual timestamps to a common sampling grid in this invention; Figure 4 This is a schematic diagram of multi-timescale prediction using three different lead times windows in this invention. Figure 5 This is a flowchart of the candidate gimbal screening, early warning, and anti-shake process in this invention; Figure 6 This is a schematic diagram of a spatial radius coverage query scenario in this invention; Figure 7 This is a schematic diagram illustrating the verification of video footage retrieved by the monitoring terminal in this invention; Figure 8 This is a functional block diagram of a small watershed flash flood early warning system for end-to-end time series prediction according to the present invention. Detailed Implementation
[0011] 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.
[0012] Example 1: Please see Figure 1 As shown, this embodiment provides an end-to-end time series prediction method for flash flood early warning in small watersheds, including: Step S10: Collect raw measurement values from multi-source heterogeneous sensing devices deployed within the small watershed to construct a multi-source heterogeneous raw data stream with dual timestamps. Perform spatiotemporal interpolation alignment and missing term completion on the multi-source heterogeneous raw data stream with dual timestamps to generate a complete data matrix. Input the complete data matrix into the end-to-end time series prediction branch for multi-timescale rolling prediction and output the flood peak forecast result. Perform hierarchical mapping based on the flood peak forecast result to obtain a four-level forecast and warning label. Further, step S10 includes: See Figure 2 Step S11: Poll and collect the raw measurement values of the multi-source heterogeneous sensing devices deployed in the small watershed, bind a station location to each multi-source heterogeneous sensing device, and add the device timestamp generated by the local clock of the multi-source heterogeneous sensing device and the receiving timestamp stamped by the data access server to the raw measurement values, and combine them to generate a multi-source heterogeneous raw data stream with double timestamps. Specifically, multi-source heterogeneous sensing devices include four types of equipment: water level gauges, surface velocity meters, rainfall meters, and weather stations. These are used to collect instantaneous water level measurements, surface velocity measurements, cumulative rainfall measurements, and multiple meteorological measurements such as wind speed, wind direction, temperature, air pressure, and humidity, respectively. Polling acquisition refers to the data access server sequentially sending read requests to each multi-source heterogeneous sensing device according to a preset polling cycle. The preset polling cycle is determined by taking the minimum value from the factory-reported cycles of each multi-source heterogeneous sensing device as a baseline value, and then dividing it by an oversampling factor. The lower limit of the oversampling factor is set to 2 to ensure that each multi-source heterogeneous sensing device completes at least two polls within one reporting cycle, thus capturing data updates within that cycle. The upper limit of the oversampling factor is the number of times the data access server completes one round of polling for all multi-source heterogeneous sensing devices per unit time. The oversampling factor is calculated by dividing the maximum allowed concurrent requests by the number of heterogeneous multi-source sensors. The oversampling factor should be between 2 and the above-mentioned acceptable multiple. If the preset polling period is too small, the request frequency between the data access server and the heterogeneous multi-source sensors exceeds the data update frequency of the sensors, resulting in numerous duplicate reads and increasing the power supply burden on the sensors. If the preset polling period is too large, multiple updates within a single reporting cycle of the heterogeneous multi-source sensor with the shortest reporting cycle are merged into a single read, smoothing out the details of minute-level spikes. If the oversampling factor is too small, only one poll may be completed within a single reporting cycle, and the phase misalignment between the factory reporting cycle and the polling cycle of the heterogeneous multi-source sensors causes the latest original measurement value to be missed. If the oversampling factor is too large, the concurrent requests to the data access server exceed the acceptable multiple, causing request queuing, new transmission delays, and affecting the accuracy of the received timestamp. For example, when the minimum reporting cycle of the multi-source heterogeneous sensing device is 30 seconds and the oversampling factor is 3, the preset polling cycle is 10 seconds. The binding station location is determined by the construction party using differential positioning equipment to measure the geographic coordinates after the multi-source heterogeneous sensing device is installed on-site and written into the device registration table of the data access server. After registration, the geographic coordinates no longer change with the data reporting of the multi-source heterogeneous sensing device. The device timestamp is generated by the local clock inside the multi-source heterogeneous sensing device. The local clock is a low-cost crystal oscillator, which is subject to long-term offset from UTC due to temperature drift and aging. The receiving timestamp is generated by the data access server using the server clock upon receiving the raw measurement value. The server clock is synchronized with UTC via network time synchronization. The multi-source heterogeneous raw data stream with dual timestamps consists of four fields: raw measurement value, device timestamp, receiving timestamp, and station location.Simultaneously retaining the device timestamp and the receiving timestamp allows step S12 to separate the actual drift component of the measurement time from the delay component of the transmission link, avoiding the failure of the entire batch of multi-source heterogeneous raw data streams with dual timestamps due to a single clock source failure, and providing an independent time reference for step S12 to construct a first-order Kalman filter.
[0013] See also Figure 2 In step S12, extract the device timestamp and the receiving timestamp of each multi-source heterogeneous sensing device in the multi-source heterogeneous raw data stream with dual timestamps, use the difference between the device timestamp and the receiving timestamp as the state variable, and construct a first-order Kalman filter; set a common sampling grid composed of equally spaced time sequences on the time axis as the alignment reference, and use the first-order Kalman filter to perform asynchronous data spatiotemporal interpolation alignment processing on the multi-source heterogeneous raw data stream with dual timestamps oriented towards the common sampling grid, and output the aligned multi-source heterogeneous data matrix; See Figure 3 This is a schematic diagram of asynchronous alignment of dual timestamps to a common sampling grid provided in an embodiment of this application. Figure 3The diagram illustrates the device's raw data with dual timestamps, equally spaced common sampling grids, and a first-order Kalman filter. Specifically, the common sampling grid refers to the equally spaced time sequence set on the time axis in step S12. The grid spacing of the common sampling grid is determined by taking the larger of the two values: the minimum reporting cycle of each multi-source heterogeneous sensor device and the time interval corresponding to the minimum input step size allowed by the end-to-end time series prediction branch in step S14. This ensures that the grid spacing is neither too fine than the minimum input step size that the end-to-end time series prediction branch can handle, nor too coarse than the reporting granularity of the fastest multi-source heterogeneous sensor device. If the grid spacing is too small, the number of columns in the aligned multi-source heterogeneous data matrix will expand, and the training and inference latency of the end-to-end time series prediction branch will increase synchronously. If the grid spacing is too large, a minute-level surge process will only fall into 1 or 2 grid points on the common sampling grid, and the end-to-end time series prediction branch will not be able to capture the rising edge of the surge process. For example, when the minimum reporting cycle of the multi-source heterogeneous sensing device is 30 seconds and the time interval corresponding to the minimum input step size is 1 minute, the larger of the two is 1 minute, and the grid spacing is 1 minute. The state variable of the first-order Kalman filter is the difference between the device timestamp and the received timestamp, the state transition matrix is an identity matrix, and the process noise variance and observation noise variance are tuned according to the nominal drift rate of the local clock provided by the multi-source heterogeneous sensing device and the network timing error of the server clock, respectively. The asynchronous data spatiotemporal interpolation alignment method uses the received timestamp as the external reference sequence. At each sampling moment, the first-order Kalman filter updates the state variable according to the latest observation between the device timestamp and the received timestamp, predicts the actual measurement time deviation at the next sampling moment, and interpolates the original measurement value back onto the common sampling grid according to the actual measurement time deviation; see continue. Figure 3The original data from the devices are misaligned in the time series, each corresponding to its own device timestamp and receiving timestamp. In this embodiment, a first-order Kalman filter is used to accurately insert these misaligned original measurements back onto the corresponding points of a common sampling grid sequence with equal spacing. A fixed grid spacing is maintained between adjacent grid points, thereby unifying the asynchronously acquired multi-source heterogeneous sensor data to the same time base. When a multi-source heterogeneous sensor has no landing point on a grid point of the common sampling grid, the predicted value of the state variable is combined with the most recent original measurement value to fill the grid point. When the difference between the device timestamp and the receiving timestamp monotonically increases over multiple consecutive sampling periods and the increment is greater than a preset jump threshold, it is determined that the local clock of the multi-source heterogeneous sensor has jumped. The state variable of the first-order Kalman filter is reset. After the reset, the state variable takes the difference between the latest device timestamp and the receiving timestamp. The process noise variance is temporarily amplified to accelerate reconvergence and avoid the jump value from polluting subsequent predictions. The preset transition threshold is determined as follows: During the manufacturing process of multi-source heterogeneous sensing devices, the nominal drift rate of the local clock is measured using a constant temperature chamber. The nominal drift rate of the local clock represents the relative deviation accumulated by the local clock relative to Coordinated Universal Time (UTC) per unit time. The product of the nominal drift rate and the accumulated duration within the continuous decision window is taken as the upper limit of the difference caused by the normal drift of the local clock within the continuous decision window and used as a benchmark. The network timing error of the server clock is then added to obtain the threshold. If the preset transition threshold is too small, normal drift fluctuations are misjudged as transitions, and the frequent reset of the first-order Kalman filter causes the already converged state variables to be repeatedly discarded. If the preset transition threshold is too large, the real local clock transitions are treated as gradual accumulations in the state variables, and the original measurement values back-interpolated onto the common sampling grid deviate from the real measurement time for a long period of time. For example, when the nominal drift rate of the local clock is 10ppm (ppm is parts per million, 10ppm means a cumulative deviation of 10 microseconds per second), the continuous decision window is 5 sampling periods, the single sampling period is the grid spacing of the common sampling grid (1 minute), the cumulative duration of the continuous decision window is 300 seconds, and the network timing error of the server clock is 20ms, the preset jump threshold is 23ms. The difference between the device timestamp and the received timestamp is an approximately linear process driven by clock drift within a short time window, which conforms to the linear Gaussian assumption of Kalman filtering. Compared with traditional linear interpolation, it more robustly restores the true measurement time. The alignment error is compressed within the range limited by the Kalman gain, avoiding the spurious fluctuations generated by traditional linear interpolation when the link jitter occurs. This provides a strictly time-aligned multi-source heterogeneous data matrix for step S13 to construct the spatial correlation diagram of multi-source heterogeneous sensing devices.
[0014] Step S13: Using the location of the monitoring station bound to the multi-source heterogeneous sensing device as graph nodes, determine the edge weights of the graph based on the geographical distance between the monitoring station locations, and construct a spatial correlation graph of the multi-source heterogeneous sensing device; locate the missing items in the aligned multi-source heterogeneous data matrix, input the aligned multi-source heterogeneous data matrix into the graph neural network driven by the spatial correlation graph of the multi-source heterogeneous sensing device, perform feature reconstruction on the missing items in the aligned multi-source heterogeneous data matrix, and generate a complete data matrix; Specifically, the graph nodes of the spatial correlation graph of multi-source heterogeneous sensing devices are taken from the set of station locations obtained after deduplication of the station locations bound to all multi-source heterogeneous sensing devices within the small watershed. The feature vector of the graph node is composed of the values of the original measurement values of all multi-source heterogeneous sensing devices bound to the corresponding station locations in the aligned multi-source heterogeneous data matrix, which are concatenated on multiple consecutive common sampling grid points covered within the preset time window length of the graph neural network. The method for determining the graph edge weight is: take the reciprocal of the geographical distance between two station locations and then take the reciprocal of the distance between all station location pairs within the small watershed. The maximum value of the number is normalized and used as the distance term, with values ranging from 0 to 1. The absolute value of the Pearson correlation coefficient of the multi-source heterogeneous data matrix aligned at two station locations within a complete historical time period is taken and truncated to between 0 and 1, serving as the correlation term. The distance term is multiplied by the correlation term to obtain the graph edge weights, with values ranging from 0 to 1. When two station locations overlap, the reciprocal of the distance is meaningless, and the corresponding graph edge is not included in the construction. The closer the geographical distance and the higher the historical correlation coefficient (after taking the absolute value), the greater the graph edge weight, reflecting the indicative significance of the upstream station location to the current station location. A complete historical time period refers to the set of continuous time periods accumulated after all multi-source heterogeneous sensing devices in the small watershed have been put into operation, where each multi-source heterogeneous sensing device simultaneously has valid original measurement values and the data integrity has been verified through the data access server. The complete historical time period is updated once before each graph neural network retraining. The method for determining the preset time window length of the graph neural network is as follows: the lag time between the pair of stations with the longest confluence time in the small watershed is taken as the lower limit, and the time occupied by a typical rainfall event in the small watershed from the start of runoff to the end of the response at the outlet section is taken as the upper limit. The values are aligned upwards by integer multiples of the grid spacing between the two. The lower limit ensures that the message transmission of the graph neural network can span the confluence process, and the upper limit ensures that the time window does not exceed the physical cycle of a typical rainfall event and avoids irrelevant historical hydrological events from being included in the feature vector. Alignment by integer multiples of the grid spacing ensures that the time window is neatly positioned on the common sampling grid. If the preset time window length of the graph neural network is too small, the downstream station has not yet perceived the hydrological event that occurred at the upstream station, and the message transmission of the graph neural network cannot span the confluence process. If the preset time window length of the graph neural network is too large, the length of the feature vector within the time window increases, and the training data requirements and inference latency of the graph neural network increase simultaneously. For example, when the lag time between the pair of stations with the longest runoff time in the small watershed is 40 minutes, the time taken for a typical rainfall event in the small watershed from the start of runoff generation to the end of the response at the outlet section is 90 minutes, and the grid spacing is 1 minute, the grid is aligned in multiples of 1 minute between 40 minutes and 90 minutes, and the preset time window length of the graph neural network is 60 minutes.The graph neural network consists of a message passing layer, a feature aggregation layer, and a missing item reconstruction layer. The message passing layer performs weighted aggregation of the feature vectors of each graph node according to the graph edge weights. The feature aggregation layer stacks multiple message passing layers to expand the receptive field of the graph nodes to multi-hop neighbors. The missing item reconstruction layer maps the feature vectors of graph nodes with missing items back to the value space of the original measurement, outputs the completed measurement value, and writes it into the complete data matrix. The training of the graph neural network uses an aligned multi-source heterogeneous data matrix within a complete historical time period as supervision. During the training phase, non-missing items in the aligned multi-source heterogeneous data matrix with a predetermined occlusion ratio are manually masked as pseudo-missing items. The original measurement value before masking is used as the label, and the loss function is the sum of the reconstructed value and the label. The mean square error between them; the method for determining the preset masking ratio is as follows: take the average proportion of naturally occurring missing items in the complete historical period of multi-source heterogeneous sensing devices in the small watershed as the lower limit, take twice the average proportion of naturally occurring missing items as the upper limit, and take a value between the two to make the distribution of pseudo-missing items in the training phase similar in magnitude to the distribution of naturally occurring missing items in the running phase, avoiding significant deviations between the training distribution and the running distribution; if the preset masking ratio is too small, the number of pseudo-missing items is insufficient, and the missing item reconstruction layer of the graph neural network cannot be adequately supervised; if the preset masking ratio is too large, there are too few non-missing items that can be referenced in the aligned multi-source heterogeneous data matrix after masking, and the message passing layer of the graph neural network cannot stably aggregate neighbor information. For example, when the average proportion of naturally occurring missing items in the complete historical period of multi-source heterogeneous sensing devices in the small watershed is 5%, the preset masking ratio is 8%.
[0015] When all upstream stations at a given location become invalid, the graph neural network regresses to a weighted combination of the historical average and the inverted value of the downstream station location as the completed measurement value. The historical average is obtained by taking the arithmetic mean of the original measurements of the corresponding station locations at the same time in the same month within a complete historical period. The inverted value of the downstream station location is calculated by combining the lag regression relationship obtained by pairing the original measurements of the downstream station location and the invalid upstream station location within a complete historical period with the convergence time as the lag, and combining this with the most recent original measurement of the downstream station location. The lag is the convergence time from the invalid upstream station location to the downstream station location, and the weighting coefficient is based on the historical average. The prediction error variances of the segment mean and the downstream station location inversion value are inversely proportionally distributed. The prediction error variance of the historical mean for the same period is directly obtained by statistically analyzing the variance between the historical mean for the same period and the original measurement value of the corresponding station location at the same moment within the complete historical time period. The prediction error variance of the downstream station location inversion value is directly obtained by statistically analyzing the variance between the downstream station location inversion value calculated according to the lag regression relationship within the complete historical time period and the original measurement value of the corresponding upstream station location at the same moment. The inverse proportional distribution means that the weighting coefficients are taken as the reciprocal of their respective prediction error variances and then normalized by the sum of the two reciprocals to make the sum of the weighting coefficients equal to 1. The smaller the prediction error variance, the larger the corresponding weighting coefficient. When the duration of the station location failure exceeds the preset tolerance time limit, the station location is temporarily removed from the spatial correlation graph of the multi-source heterogeneous sensing equipment, and the graph edge weights adjacent to the station location are recalculated to prevent noise from the failed station location from propagating to adjacent station locations. The method for determining the preset tolerance time limit is as follows: take an integer multiple of the factory reporting cycle of the multi-source heterogeneous sensing equipment as the base time, and take the typical continuous packet loss frame number of the historical sporadic short-term packet loss of the multi-source heterogeneous sensing equipment in the small watershed as the integer multiple, so that the base time can cover the sporadic short-term packet loss process without judging sporadic packet loss as continuous failure; and then add the on-site maintenance arrival time. If the preset tolerance time limit is too small, the sporadic short-term packet loss of the multi-source heterogeneous sensing equipment will be frequently judged as continuous failure, and the frequent reconstruction of the spatial correlation map of the multi-source heterogeneous sensing equipment will cause the source of the completed measurement value to switch repeatedly between adjacent time points. If the preset tolerance time limit is too large, the multi-source heterogeneous sensing equipment that has been physically damaged will be hung on the spatial correlation map of the multi-source heterogeneous sensing equipment for a long time, and the erroneous map node will continue to propagate noise to the location of adjacent stations. For example, when the factory reporting cycle of multi-source heterogeneous sensing devices is 60s, the typical number of consecutive packet loss frames of historical sporadic short-term packet loss of multi-source heterogeneous sensing devices in a small watershed is 15, the basic duration is 15 times the reporting cycle, resulting in 15 minutes, and the on-site maintenance arrival time is 30 minutes, the preset tolerance time limit is set to 45 minutes.The measurement values are supplemented by the physical sources of adjacent stations to support the missing items. The failure of a single multi-source heterogeneous sensing device no longer causes the overall data link to be interrupted, providing missing and comparable input for the multi-timescale rolling prediction in step S14.
[0016] Step S14: Truncate the complete data matrix to the forecast window length, construct multiple sets of input tensors, push the multiple sets of input tensors into the corresponding end-to-end time series prediction branches to perform rolling updates, and integrate the forecast values returned by each end-to-end time series prediction branch to output the flood peak forecast result; the flood peak forecast result includes the flood peak water level forecast value, surface flow velocity forecast value, and flood peak occurrence time forecast value; See Figure 4 This is a schematic diagram of multi-timescale prediction with three forecast windows provided in an embodiment of this application. Figure 4 The mid-timeline illustrates the historical and forecast periods divided by the current moment, as well as three forecast windows: short-term, medium-term, and long-term. Combined with... Figure 4 As shown, specifically, end-to-end time series forecasting branches are divided into three types according to the forecast window length: end-to-end time series forecasting branches with short forecast window lengths, end-to-end time series forecasting branches with medium forecast window lengths, and end-to-end time series forecasting branches with long forecast window lengths. See also... Figure 4These three end-to-end time series prediction branches are responsible for capturing minute-level surge processes, characterizing the runoff generation and confluence processes in small watersheds, and incorporating external forcing of small watersheds by meteorological forecasts, respectively, and integrating and outputting peak flood forecast results within the prediction period. The methods for determining the length of the three lead time windows are as follows: the short lead time window length is taken as the shortest observation time between the initial rise and the warning water level threshold at the location of the fastest-responding monitoring station within the small watershed; the medium lead time window length is taken as the runoff generation and confluence time at the outlet section of the small watershed; and the long lead time window length is taken as the lower limit of the effective forecast duration provided by the meteorological forecast source. If the lead time window length is too small, the corresponding end-to-end time series prediction branch cannot cover the time scale of the physical process to be predicted, resulting in insufficient lead time for the peak flood level forecast; if the lead time window length is too large, the corresponding end-to-end time series prediction branch requires longer historical input and reduces the available training sample size, increasing the variance of the peak flood level forecast. For example, when the shortest observation time between the initial rise and the warning water level threshold at the fastest response monitoring station is 30 minutes, the runoff generation and confluence time at the small watershed outlet section is 2 hours, and the lower limit of the effective forecast duration provided by the meteorological forecast source is 12 hours, the short forecast window length is 30 minutes, the medium forecast window length is 2 hours, and the long forecast window length is 12 hours. The end-to-end time series prediction branch corresponding to each forecast window length uses a slice of the complete data matrix within the corresponding forecast window length as the input tensor. The number of columns in the input tensor is obtained by dividing the forecast window length by the grid spacing, and the number of rows is obtained by the number of channels in the complete data matrix. The end-to-end time series prediction branch corresponding to each forecast window length consists of an encoder, a temporal memory module, and a multi-task output head. The encoder maps the input tensor to the hidden state space. The temporal memory module performs state retention and forgetting in the time dimension for the hidden states. The temporal memory module is implemented using a gated loop structure such as a long short-term memory network or a gated loop unit. The gated loop structure uses a forget gate to gradually decay components in the historical hidden state that are no longer relevant to the latest complete data matrix according to the gate weights, and an update gate to write new information carried in the latest complete data matrix into the hidden state according to the gate weights. This preserves the near-term memory of minute-level surge processes and the far-term memory of small-basin runoff generation processes in the hidden state. The multi-task output head outputs the flood peak water level forecast, surface velocity forecast, and flood peak occurrence time forecast, respectively. The end-to-end time series prediction branches corresponding to the three forecast window lengths share the target structure of the multi-task output head, but their parameters are independent, avoiding mutual interference between time scales caused by parameter sharing. A rolling prediction is triggered every preset refresh interval, pushing the latest frame of the complete data matrix into the end-to-end time series prediction branches corresponding to the three forecast window lengths, and discarding the oldest frame in the input tensor of each end-to-end time series prediction branch, resulting in a rolling updated flood peak forecast.The preset refresh interval is determined as follows: the grid spacing is taken as the lower limit, and the requirement for forecast refresh frequency in the small watershed emergency response scheme is taken as the upper limit, with a value between the two. If the preset refresh interval is too small, the inference load of the end-to-end time series prediction branch increases, and the difference between consecutive frames of flood peak forecast results is too small, causing excessive triggering of downstream modules. If the preset refresh interval is too large, the response of the flood peak forecast results to the latest complete data matrix is delayed, and minute-level surge processes are missed between two refreshes. For example, when the grid spacing is 1 minute and the requirement for forecast refresh frequency in the small watershed emergency response scheme is no less than once every 5 minutes, the preset refresh interval is 5 minutes. When the difference between the peak flood level forecasts output by the end-to-end time series forecast branch corresponding to the short-forecast window length and the end-to-end time series forecast branch corresponding to the medium-forecast window length exceeds a preset branch consistency threshold, the peak flood level forecasts output by the end-to-end time series forecast branch corresponding to the long-forecast window length are subject to weight reduction fusion. Weight reduction fusion is achieved by linearly reducing the output weights of the end-to-end time series forecast branch corresponding to the long-forecast window length according to the normalized value of the difference between the peak flood level forecasts output by the end-to-end time series forecast branches corresponding to the short-forecast window length and the medium-forecast window length, thus avoiding long-term external forcing dominating short-term urgent warnings. The preset branch consistency threshold is determined as follows: One part of the difference between the warning water level threshold and the immediate evacuation water level threshold within the flood control water level threshold in the small watershed is divided into three equal parts as a benchmark. This benchmark is then superimposed with the root mean square error of the end-to-end time series prediction branch on the historical validation set. Dividing the threshold into three equal parts corresponds to the water level granularity of each level after evenly distributing the overall span from the warning water level threshold to the immediate evacuation water level threshold according to three levels (warning, preparation for evacuation, and immediate evacuation). This ensures that the branch consistency threshold and the granularity of each level of the flood control water level threshold are on the same order of magnitude, preventing the deviation between branch outputs from being absorbed by the three levels of the flood control water level threshold. If the preset branch consistency threshold is too small, conventional model noise is misjudged as branch inconsistency, and the output weight of the end-to-end time series prediction branch corresponding to the long-term forecast window length is repeatedly downgraded. If the preset branch consistency threshold is too large, conflicts between actual time scales do not trigger weighted fusion, and long-term external pressure dominates short-term urgent warnings. For example, when the difference between the warning water level threshold and the immediate evacuation water level threshold in the flood control water level threshold is 0.6m, and the root mean square error of the end-to-end time series prediction branch on the historical validation set is 0.1m, one part after being divided into 3 equal parts is 0.2m, and then the 0.1m root mean square error is added, and the preset branch consistency threshold is 0.3m.The historical validation set is a subset of the complete historical time period that is not involved in the training of the graph neural network and the training of the end-to-end time series prediction branch, and is used to independently estimate the missing term reconstruction error of the graph neural network and the flood peak water level forecast error of the end-to-end time series prediction branch.
[0017] Step S15: Perform hierarchical mapping based on the comparison between the flood peak water level forecast value in the flood peak forecast result and the flood limit water level threshold pre-configured at the station location to generate the corresponding four-level forecast and early warning identifier; the flood limit water level threshold includes the early warning water level threshold, the prepared evacuation water level threshold, and the immediate evacuation water level threshold.
[0018] Specifically, the flood control water level thresholds are pre-configured by the flood control authorities of small watersheds according to the location of monitoring stations. The relationship between the three flood control water level thresholds is as follows: the immediate evacuation water level threshold is greater than the prepared evacuation water level threshold, which is greater than the warning water level threshold. The warning water level threshold represents the starting point at which the water level measurement at the monitoring station exceeds the daily fluctuation and enters the risk concern zone; the prepared evacuation water level threshold represents the starting point at which personnel in the risk area need to begin assembling and supplies need to be in place; and the immediate evacuation water level threshold represents the starting point at which personnel in the risk area need to be evacuated immediately. The judgment logic for the graded mapping is as follows: when the flood peak water level forecast value in the flood peak forecast result is greater than or equal to the immediate evacuation water level threshold, it is mapped to a red warning; when the flood peak water level forecast value is greater than or equal to the prepared evacuation water level threshold but less than the immediate evacuation water level threshold, it is mapped to an orange warning; when the flood peak water level forecast value is greater than or equal to the warning water level threshold but less than the prepared evacuation water level threshold, it is mapped to a yellow warning; and when the flood peak water level forecast value is less than the warning water level threshold, it is mapped to a blue warning. The mapping results are output in the form of a four-level forecast and early warning identifier. The four-level forecast and early warning identifier carries the early warning level, station location, flood peak water level forecast, surface velocity forecast, and flood peak occurrence time forecast. The early warning level is taken from one of the red, orange, yellow, and blue early warnings obtained from the hierarchical mapping judgment logic. It provides a complete event context to the video verification stage in step S20. The four-level forecast and early warning identifier is fully aligned with the four-level early warning system of the small watershed monitoring and early warning and full-cycle event supervision system. It can drive the hierarchical preemptive PTZ scheduling in step S20 without further mapping conversion. The small watershed monitoring and early warning and full-cycle event supervision system refers to the field business system that receives the four-level forecast and early warning identifier, dispatches the duty terminal to conduct video verification, and archives the early warning handling work order. It includes a terminal display module for video verification, a flood history record database for work order archiving, and a post-disaster review and analysis module for post-disaster assessment.
[0019] Flash flood events in small watersheds are driven by both rainfall pulses and meteorological external forcing. Water level, surface velocity, rainfall, and meteorological measurements from multiple sources need to be simultaneously input into the prediction model to provide meaningful peak water level forecasts. However, the sampling frequencies, reporting cycles, and link delays of water level gauges, surface velocity meters, rainfall meters, and meteorological stations are all different, leading to misalignment and drift in the timestamps of the original measurements. Occasional failures of a single multi-source heterogeneous sensor can drag down the entire data link. The conventional approach is to linearly interpolate the original measurements from each multi-source heterogeneous sensor at a uniform sampling interval and then directly input them into a time series model with a single window length. When the link jitters, the interpolation produces spurious fluctuations that are amplified by the model as true signals. Minute-level spikes and hourly-level gradual changes mask each other. When a single multi-source heterogeneous sensor fails, the entire series of original measurements is filled with the mean or zeros, resulting in frequent false alarms from the model. Step S10 preserves an independent benchmark for delay correction and clock jump detection using multi-source heterogeneous raw data streams with dual timestamps. This data is then strictly aligned on a common sampling grid using a first-order Kalman filter, avoiding spurious fluctuations caused by link jitter in traditional linear interpolation. A graph neural network driven by the spatial correlation graph of the multi-source heterogeneous sensing devices performs feature reconstruction on missing items, ensuring that a single multi-source heterogeneous sensing device failure no longer interrupts the overall data link. The completed measurements have an interpretable physical source from adjacent station locations. End-to-end time series prediction branches corresponding to three different lead time window lengths respectively carry minute-level surges, runoff generation processes, and meteorological external forcing, avoiding mutual masking between time scales caused by a single window length. Rolling prediction refresh and branch consistency threshold weighting fusion prevent long-term external forcing from dominating short-term urgent warnings. A four-level forecast warning identifier is output through graded mapping of flood limit water level thresholds, directly connecting to the video verification step in S20. Therefore, step S10 transforms the technical problems of misalignment, incompleteness, and inaccuracy in multi-source heterogeneous sensor data into flood peak forecast results that are strictly aligned in the time dimension, physically interpretable in the spatial dimension, and physically separable in the time scale. It is noteworthy that the edge weights of the spatial correlation graph of the multi-source heterogeneous sensor devices integrate geographical distance and historical correlation coefficients. When some upstream monitoring stations enter the concentrated rainstorm phase of the high-water season, the historical correlation coefficients adaptively increase the weight of the main rainstorm control station's location in the graph neural network message transmission. The physical source of the completed measurement values adjusts according to the current hydrological situation. Compared to conventional static spatial interpolation, which uses the same set of weights evenly in both the high-water and low-water seasons, this provides a more accurate completion result for the high-water season, closely reflecting the actual confluence. This situational adaptive characteristic is often underestimated before the system is put into operation, but it becomes a key contribution to supporting the accuracy of high-level early warnings after operation.
[0020] Step S20: For the Level 4 forecast and early warning identifier, perform spatial radius coverage query and time dimension anti-shake filtering to obtain the candidate camera PTZ resource ranking list and the effective early warning event sequence. Based on the candidate camera PTZ resource ranking list and the effective early warning event sequence, perform hierarchical preemptive PTZ lock signaling interaction to determine the target camera PTZ. Establish a concurrent request conflict queue and output the conflict arbitration result to control the duty terminal to pull the video image from the target camera PTZ. Bind the video image with the effective early warning event sequence and the flood peak forecast result to generate an early warning handling work order.
[0021] Further, step S20 includes: See Figure 5 Step S21: For each level 4 forecast and early warning sign associated with a station location, perform a radius coverage query to identify candidate camera PTZ resources, and sort the candidate camera PTZ resources in ascending order of geographical distance from the station location to generate a candidate camera PTZ resource sorting list. See Figure 6 This is a schematic diagram of a spatial radius coverage query scenario provided in an embodiment of this application. Figure 6 The diagram illustrates the location of the measuring station, the circular area defined by the video verification radius, the candidate camera gimbal within the circle, and the camera gimbal outside the circle. Combined with... Figure 6 As shown, specifically, radius coverage query refers to selecting camera pan-tilt units (PTZ units) deployed within a small watershed that are geographically within a distance of less than or equal to the preset video verification radius and whose field of view orientation can include the station location in the imaging field of view, using the station location as the center and a preset video verification radius as the query radius. These PTG units are then considered as candidate camera pan-tilt unit resources. Figure 6In this scenario, Candidates 1, 2, and 3 are all located within the video verification radius, and their field of view includes the water level gauge at the station location, thus they are designated as candidate resources. Devices located at a distance greater than the preset radius are excluded, as indicated by "not selected outside the circle," meaning "cameras with pan-tilt units outside the radius" are directly eliminated. Whether the field of view covers the station location is determined based on the field of view parameters provided by the camera pan-tilt unit at the factory and the installation azimuth parameters calibrated on-site. The preset video verification radius is determined as follows: the effective imaging distance provided by the camera pan-tilt unit at the factory is taken as the upper limit reference, and the farthest observation distance required to distinguish the water level scale under the sensor resolution of the camera pan-tilt unit is taken as the lower limit reference. The value between the two, including the endpoints, is taken. The value closer to the lower limit prioritizes ensuring the ability of the image to distinguish the water level scale, and the value closer to the upper limit prioritizes ensuring the coverage of the candidate camera pan-tilt unit resources. If the preset video verification radius is too small, there are no camera pan-tilt units available around the location of some level 4 forecast and warning signs, the candidate camera pan-tilt unit resource sorting list is empty, and step S23 cannot initiate the locking request. If the preset video verification radius is too large, camera pan-tilt units that are far away from the location of the monitoring station are included in the candidate camera pan-tilt unit resource sorting list, and the water level scale is too small in the image to support the video verification and interpretation of the monitoring terminal. For example, when the effective imaging distance of the camera pan-tilt unit is 5km and the maximum observation distance required to distinguish water level scale markings is 3km, priority is given to ensuring the ability to distinguish water level scale markings in the image, and the preset video verification radius is 3km. The geographical distance is calculated by converting the geographical coordinates of the station location and the camera pan-tilt unit into the great circle distance. The candidate camera pan-tilt unit resources are rearranged in ascending order of geographical distance to obtain a candidate camera pan-tilt unit resource ranking list. The camera pan-tilt unit at the top of the candidate camera pan-tilt unit resource ranking list corresponds to the camera pan-tilt unit closest to the station location and with the strongest image resolution capability. The camera pan-tilt units in the next and later positions serve as alternatives when the first camera pan-tilt unit is occupied. Step S21 completes the spatial dimension candidate screening and sorting in one go when the Level 4 forecast and warning sign arrives, avoiding the runtime delay caused by the real-time geographic query in the locking request stage of step S23. This compresses the overall response time of the hierarchical preemptive PTZ lock signaling interaction within the budget stipulated by the small watershed emergency response plan, and provides an orderly attempt sequence for step S23 to attempt locking requests in order of geographic distance from near to far. If step S21 is missing, each locking request in step S23 must calculate the geographic distance and field of view orientation coverage relationship on-site. The initiation delay of the locking request increases linearly with the scale of candidate camera PTZ resources, and the video verification channel for urgent warnings cannot be opened within the budgeted time.
[0022] See also Figure 5Step S22: Within the preset anti-shake time window, slide count the Level 4 forecast and warning icons, filter out the Level 4 forecast and warning icons that have not reached the effective counting threshold, and summarize the Level 4 forecast and warning icons that have reached the effective counting threshold into a valid warning event sequence. Specifically, the preset anti-shake time window length is used to perform a limited-length review of the Level 4 forecast and warning signs along the time axis. The sliding count refers to counting the cumulative number of Level 4 forecast and warning signs of the same warning level at the same station location within the preset anti-shake time window length when each frame of the Level 4 forecast and warning sign arrives. The warning level is taken from one of the red warning, orange warning, yellow warning, and blue warning generated in step S15. The method for determining the preset anti-shake time window length is as follows: take the preset refresh interval mentioned in step S14 multiplied by an integer multiple required to ensure that at least 3 frames of level 4 forecast warning labels appear within the preset anti-shake time window length as the lower limit reference, and take the maximum waiting time allowed by the small watershed emergency response plan for each warning level as the upper limit reference. Between the two, the value is increased from red warning to blue warning according to the warning level. The higher the warning level, the shorter the preset anti-shake time window length. If the preset anti-shake time window length is too small, the number of observable frames within the preset anti-shake time window length is insufficient, and the isolated level 4 forecast warning labels generated by a single model noise crossing are treated as valid warning events and added to the valid warning event sequence, and the downstream locking request stage is frequently triggered. If the preset anti-shake time window length is too large, the real warning transition needs to accumulate enough frames over a longer period of time before it is released, and the urgent warning of the minute-level surge process is delayed. For example, when the preset refresh interval is 5 minutes and the maximum allowed waiting time for the small watershed emergency response plan to respond to the blue warning is 60 minutes, the preset anti-shake time window length is 15 minutes for the red warning, 20 minutes for the orange warning, 30 minutes for the yellow warning, and 45 minutes for the blue warning. The effective counting threshold is determined as follows: half of the integer part obtained by dividing the preset anti-shake time window length by the preset refresh interval and rounding down is taken as the baseline, and then the average number of times the flood peak water level forecast value caused by model noise crosses the flood limit water level threshold is added. Half corresponds to the judgment logic of "at least half of the observation frames in the window support the same warning level before release", so that the effective counting threshold expands and contracts proportionally with the anti-shake time window length. If the effective counting threshold is too small, a small number of model noise crossings will be judged as effective warning events, and invalid preemption requests generated near the threshold boundary cannot be suppressed at the source. If the effective counting threshold is too large, the real warning transition is also difficult to reach the judgment threshold, the effective warning event sequence will idle for a long time, and the downstream locking request stage will not be triggered for a long time. For example, when the preset anti-shake time window length is 30 minutes for a yellow warning, the preset refresh interval is 5 minutes, and the average number of times the flood peak water level forecast value caused by model noise crosses the flood limit water level threshold is 1, 30 minutes divided by 5 minutes gives an integer part of 6, half of which is rounded down to 3, and then one model noise crossing is added, the effective count threshold is 4.When the highest warning level of the Level 4 forecast warning indicator within the preset anti-shake time window is higher than the warning level of the most recently added valid warning event in the valid warning event sequence, the immediate release rule is triggered. The immediate release rule skips the waiting time for the accumulated sliding count and directly adds the Level 4 forecast warning indicator with the highest warning level as a valid warning event to the valid warning event sequence. The immediate release rule enables the preset anti-shake time window to filter out threshold boundary fluctuations for blue, yellow, and orange warnings, without delaying urgent warnings such as red warnings. The valid warning event sequence entering the hierarchical preemptive gimbal lock signaling interaction is all real warning transitions. If step S22 is missing, the locking request stage in step S23 will be repeatedly triggered by the instantaneous Level 4 forecast warning indicator generated by the model noise crossing the flood limit water level threshold boundary. The camera gimbal orientation will repeatedly switch between the two triggers, causing mechanical damage and unstable video images.
[0023] Step S23: Maintain a hierarchical preemptive PTZ lock for each camera PTZ. Based on the warning level carried by the effective warning event sequence, initiate a locking request to each camera PTZ along the candidate camera PTZ resource sorting list to perform hierarchical preemptive PTZ lock signaling interaction, determine the PTZ lock occupancy status of each camera PTZ and authorize the locking request to lock the target camera PTZ. Specifically, the target camera PTZ refers to the camera PTZ currently being attempted to be locked or already locked by a locking request during the hierarchical preemptive PTZ lock signaling interaction process. As the locking request is initiated sequentially along the candidate camera PTZ resource sorting list, the target camera PTZ points to the corresponding camera PTZ in the list as the attempt progresses. The hierarchical preemptive PTZ lock is a separate status recording unit maintained for each camera PTZ. It records four items: the current lock-holding warning level, the lock-holding terminal identifier, the lock-holding start time, and the maximum lock-holding duration. The lock-holding terminal identifier points to the currently holding terminal. The terminal is a human-machine interface terminal with video verification and feedback permissions in the small watershed monitoring and early warning and full-cycle event supervision system. Terminals are categorized into three levels based on the duty officer: grassroots terminal, town-level terminal, and county-level terminal, with the level identifier increasing sequentially. The PTZ lock occupancy status has two values: idle and occupied. When the current warning level of the locked device is empty, the PTZ lock occupancy status is idle; when the current warning level of the locked device is one of the red, orange, yellow, or blue warnings, the PTZ lock occupancy status is occupied. The method for determining the maximum maximum holding time is as follows: the median time required for the monitoring terminal to complete one video verification and fill in the handling feedback is taken as the lower limit reference, and the maximum allowable time for a single camera PTZ to be occupied in the small watershed emergency response plan is taken as the upper limit reference. The value decreases from red warning to blue warning, and the higher the warning level, the longer the maximum maximum holding time. If the maximum maximum holding time is too small, the monitoring terminal will be forcibly released before completing the video verification, resulting in frequent screen switching and the inability to form a closed loop for handling feedback. If the maximum maximum holding time is too large, the monitoring terminal that has completed the video verification will occupy the camera PTZ for a long time, and newly arrived valid warning events of the same or lower level will not be responded to for a long time. For example, when the median time required for a monitoring terminal to complete one video verification and fill in the feedback is 3 minutes, and the maximum allowed time for a single camera pan-tilt unit to be occupied under the small watershed emergency response plan is 15 minutes, the maximum lock holding time is 15 minutes for red alerts, 10 minutes for orange alerts, 6 minutes for yellow alerts, and 3 minutes for blue alerts. The lock request is triggered by a valid alert event in the valid alert event sequence. The lock request carries five items: the location of the monitoring station corresponding to the valid alert event, the alert level of the valid alert event, the identifier of the monitoring terminal initiating the lock request, the level identifier of the monitoring terminal initiating the lock request, and the timestamp of the valid alert event.The hierarchical preemptive PTZ lock signaling interaction initiates lock requests sequentially from the first in the candidate camera PTZ resource sorting list to the last: When the target camera PTZ lock is idle, a lock request is granted, the current lock-holding warning level of the hierarchical preemptive PTZ lock is updated to the warning level carried in the lock request, the lock-holding terminal identifier is updated to the identifier of the monitoring terminal that initiated the lock request, the lock-holding start time is updated to the current server clock time, and a PTZ orientation switching command is issued to make the target camera PTZ align with the test station position; when the target camera PTZ lock is occupied and the current lock-holding warning level is lower than the warning level carried in the lock request, a lock release command is issued to the monitoring terminal pointed to by the lock-holding terminal identifier, the lock release command carries the newly arrived warning level of the lock request and the test station position, and waits for the lock-holding terminal to finish holding the lock. After the terminal indicated by the terminal identifier confirms the receipt, the locking request is granted. When the target camera's PTZ lock is occupied and the current lock's warning level is greater than or equal to the warning level carried in the locking request, the target camera PTZ is skipped, and a locking request is initiated to the next best camera PTZ in the candidate camera PTZ resource sorting list. This process is repeated sequentially along the candidate camera PTZ resource sorting list until an authorizable camera PTZ is found. When all camera PTZs in the candidate camera PTZ resource sorting list are occupied and the current lock's warning level is greater than or equal to the warning level carried in the locking request, the locking request is added to the concurrent request conflict queue of the first camera PTZ in the candidate camera PTZ resource sorting list, awaiting subsequent authorization according to the lock handover management in step S24. The warning levels are determined in descending order: red warning, orange warning, yellow warning, and blue warning. Instead of forcibly interrupting the currently holding terminal, the acknowledgment of the lock command is used. The camera pan-tilt motor will not be interrupted by command conflicts and the image switching has a predictable transition time, avoiding the motor from jamming or shaking when command conflicts occur. Low-priority lock requests are redirected to the suboptimal camera pan-tilt. Low-alert-level requests can still obtain video verification channels without crowding out the physical resources of high-alert-level requests. High-alert-level requests can still obtain the camera pan-tilt that is closest to the station and has the strongest image resolution when physical resources are scarce.
[0024] Step S24: Collect multiple locking requests that arrive at the target camera PTZ at the same time to establish a concurrent request conflict queue. Perform multi-dimensional priority sorting and lock handover management on multiple locking requests in the concurrent request conflict queue according to the PTZ lock occupancy status, and output the conflict arbitration result for multiple locking requests. Specifically, the concurrent request conflict queue is a priority queue maintained separately for each target camera PTZ, sorted by multiple dimensions. Multiple locking requests arriving at the target camera PTZ at the same time are sorted by multiple dimensions, with the priority sorting based on three key comparisons: the first priority is the warning level carried by the locking request, decreasing from red, orange, yellow, to blue; the second priority is the timestamp of the valid warning event corresponding to the locking request, with earlier timestamps having higher priority; the third priority is the level identifier of the terminal initiating the locking request, with higher level identifiers having higher priority. When the first priority is the same, the second priority is used for comparison; when both the first and second priorities are the same, the third priority is used for comparison, resulting in the locking request at the head of the concurrent request conflict queue. The lock handover process, step S24, is triggered when the target camera PTZ lock is released or released from its PTZ lock occupancy state: the pending locking request is retrieved from the head of the concurrent request conflict queue and granted lock according to the authorization process in step S23. After completion, the granted locking request is dequeued. When a new locking request with the same warning level as the currently held lock exists in the concurrent request conflict queue, the remaining locking time is calculated using the lock start time and the maximum maximum locking duration. The remaining locking time is obtained by subtracting the difference between the current server clock time and the lock start time from the maximum maximum locking duration. The lock handover of the head locking request is only executed after the remaining locking time reaches zero, avoiding frequent switching of locking requests with the same warning level that could cause repeated shaking of the camera pan / tilt direction. When the length of the concurrent request conflict queue exceeds the preset queue length limit, the low-warning-level locking requests at the tail of the concurrent request conflict queue are downgraded. Downgrading means replacing the target camera pan / tilt of the low-warning-level locking request with the second-best camera pan / tilt in the candidate camera pan / tilt resource sorting list and re-enqueuing it. After re-enqueuing, the level 3 keyword sorted according to multi-dimensional priority continues to participate in the sorting. The method for determining the preset queue length upper limit is as follows: The total number of monitored terminals within the small watershed is used as a lower limit reference, and the upper limit of the number of locking requests that may be reached within a complete locking cycle of the target camera pan-tilt unit is used as an upper limit reference; a value between these two is taken. If the preset queue length upper limit is too small, the concurrent request conflict queue will frequently overflow, causing normal locking requests to be incorrectly downgraded, and the suboptimal camera pan-tilt units in the candidate camera pan-tilt unit resource sorting list will be repeatedly occupied. If the preset queue length upper limit is too large, long queue waiting will cause locking requests near the tail of the queue to not receive a response for a long time, and effective warning events will be stuck in the concurrent request conflict queue, missing the urgent handling window. For example, when the total number of monitored terminals within the small watershed is 8 and the upper limit of the number of locking requests that may be reached within a complete locking cycle of the target camera pan-tilt unit is 12, the preset queue length upper limit is set to 10.When multiple monitored terminals simultaneously request the same target camera PTZ, the concurrent request conflict queue provides a unique conflict arbitration result. The conflict arbitration result indicates which locking request should be granted and which should be redirected to the suboptimal camera PTZ, eliminating camera PTZ orientation jitter caused by multiple monitored terminals competing for control, and providing a single, clear target camera PTZ for step S25 to pull the real-time video stream from the target camera PTZ that has obtained locking permission.
[0025] Step S25: In accordance with the conflict arbitration result, the guard terminal that has obtained the locking permission pulls the real-time video stream from the target camera pan-tilt unit, captures the video images linked with the effective early warning event sequence, binds and splices the video images with the effective early warning event sequence and the flood peak forecast results, and establishes an early warning handling work order.
[0026] See Figure 7 This is a schematic diagram of a monitoring terminal retrieving video footage for verification, provided in an embodiment of this application. Figure 7 The diagram illustrates the interaction between the target camera's pan / tilt unit, the monitoring station's location, the monitoring terminal, and the early warning and response work order. Combined with... Figure 7 As shown, specifically, the real-time video stream is output from the target camera pan-tilt unit via a streaming media server using a video streaming protocol, such as... Figure 7 In this scenario, the target camera's pan-tilt unit captures the location of the monitoring station, including the water level gauge. The monitored terminal, having obtained lock access, retrieves a real-time video stream from the streaming media server using a streaming protocol chosen from RTSP, WebRTC, or FLV, consistent with the protocol used by the real-time video function of the small watershed monitoring and early warning and full-cycle event supervision system. The video frame refers to the sequence of images displayed on the monitoring terminal's display module. Binding and splicing refers to combining the video frame, the corresponding valid early warning events in the valid early warning event sequence, the corresponding flood peak water level forecast, surface velocity forecast, and flood peak occurrence time forecast from the flood peak forecast results, along with the handling feedback submitted by the monitored terminal, into a single early warning handling work order. (See also...) Figure 7The monitoring terminal integrates the acquired video footage evidence, flood peak forecast, and response feedback by performing a binding and splicing operation, ultimately generating a complete early warning and response work order. The early warning and response work order is spliced by fields, including five categories: basic early warning fields, geographic information fields, video footage evidence fields, flood peak forecast result fields, and response feedback fields. These fields are linked through the identifier of a valid early warning event and written into the flood history database of the small watershed monitoring and early warning and full-cycle event supervision system. Response feedback refers to the on-site interpretation results submitted by the monitoring terminal to the small watershed monitoring and early warning and full-cycle event supervision system after the terminal displays the video footage and confirms the location of the monitoring station. Response feedback includes three categories of fields: hazard type, on-site photos, and textual comments. When the PTZ lock occupancy status of the target camera changes, the access address of the real-time video stream is pushed to the monitoring terminal that has obtained locking permission. After the monitoring terminal with locking permission loads the access address of the real-time video stream on its terminal display module, it pulls the real-time video stream from the streaming media server according to the video stream protocol, sees the video image of the monitoring station location, and submits a handling feedback within the preset handling time limit. After the handling feedback field of the warning handling work order is filled in, it enters the closed-loop archiving state. If the monitoring terminal fails to submit a handling feedback within the preset handling time limit, it is transferred to a monitoring terminal with a higher level identifier and the locking request is re-initiated according to the authorization process in step S23 to prevent effective warning events from being missed due to the lack of response from a single monitoring terminal. The method for determining the preset handling time limit is as follows: the median time required for the monitoring terminal to confirm the video footage and form an on-site judgment is taken as the lower limit reference, and the time limit required for the emergency response plan for small watersheds to issue personnel transfer decisions is taken as the upper limit reference. The value is incremented according to the warning level from red warning to blue warning. The higher the warning level, the shorter the preset handling time limit. If the preset handling time limit is too small, the normal video verification of the monitoring terminal will not be completed before it is transferred to the next level monitoring terminal, the next level monitoring terminal will be disturbed for no reason, and the overall dispatch link load will increase. If the preset handling time limit is too large, effective warning events that have no response for a long time will not be transferred to the next level monitoring terminal, and the urgent handling window will be missed, resulting in missed reports. For example, when the median time required for the monitoring terminal to confirm the video footage and form an on-site judgment is 3 minutes, and the time limit for issuing personnel transfer decisions in the small watershed emergency response plan is 30 minutes, the preset handling time limits are 5 minutes for red alerts, 10 minutes for orange alerts, 20 minutes for yellow alerts, and 30 minutes for blue alerts. The early warning handling work order is directly connected to the post-disaster review and analysis module of the small watershed monitoring and early warning and full-cycle event supervision system, serving as the original evidence for the post-disaster assessment report and the retrospective of basic watershed information, thus realizing a complete closed loop from the fourth-level forecast and early warning label output in step S15 to the early warning handling work order formed in step S25.
[0027] The on-site handling of flash flood warnings in small watersheds relies on video verification of the station locations associated with the Level 4 forecast warning markers using camera pan-tilt units. When the flood peak forecast triggers the Level 4 forecast warning marker, multiple monitoring terminals often simultaneously request the same camera pan-tilt unit. Furthermore, near the flood limit water level threshold, the Level 4 forecast warning marker can repeatedly cross the threshold due to model noise, generating instantaneous Level 4 forecast warning marker floods, further exacerbating the situation of multiple terminals concurrently competing for the camera pan-tilt unit. The conventional first-come, first-served scheduling method grants camera pan-tilt units according to the order of locking requests, resulting in the camera pan-tilt unit orientation being occupied by earlier low-level warning requests, while subsequent high-level warning requests have to wait, and the video verification channel at critical moments is missed. The instantaneous locking requests generated by threshold boundary jitter repeatedly switch the camera pan-tilt unit orientation, causing frequent motor operation and even damage. The handling process is already hampered by noise at the underlying locking level. Step S20 maps the Level 4 forecast and warning identifiers to the camera PTZ units via spatial radius coverage query before the arrival of the effective warning event. During the locking request phase, the candidate camera PTZ unit resources are sorted in order of geographical distance from near to far. When physical resources are scarce, the camera PTZ unit with the closest distance to the station and the strongest image resolution can still be prioritized for higher warning levels. By combining sliding count within the preset anti-shake time window with the immediate release rule, model noise crossing near the flood limit water level threshold is filtered out at the locking link entry, and urgent red warnings are not delayed by the preset anti-shake time window. The hierarchical preemptive PTZ lock records the current warning level, lock-holding terminal identifier, lock-holding start time, and longest holding time. The lock duration is capped, and a confirmation of the lock command replaces forced interruption. High-alert-level requests gain the right to grant the lock in the lock request competition without interrupting the camera pan-tilt motor due to command conflicts. Through a three-level keyword priority sorting of the concurrent request conflict queue and a downgraded processing triggered by a preset queue length limit, a unique conflict arbitration result is given when multiple terminals simultaneously request the same target camera pan-tilt. Low-alert-level lock requests are redirected to the next best camera pan-tilt without blocking the concurrent request conflict queue. The video footage, valid early warning events, flood peak forecast results, and handling feedback are completely bound and stitched together through an early warning handling work order. The closed loop from the four-level forecast and early warning identifier output in step S15 to the early warning handling work order formed in step S25 is connected. Through the cooperation of the above branches, step S20 transforms the intermediate technical problem of multiple terminals concurrently competing for the same camera pan-tilt and the impact of threshold boundary jitter floods on the underlying lock-grabbing link into an orderly scheduling process where high-alert-level requests prioritize video verification, low-alert-level requests are orderly alternatives, motor actions are not interrupted, and handling feedback is not missed.It is worth noting that the preset anti-shake time window length decreases from blue to red according to the warning level. Combined with the immediate release rule for high-alert-level bypass release, the overall emergency locking link exhibits a dual-speed behavior near the flood limit threshold: long anti-shake for low-alert-level and instantaneous penetration for high-alert-level. Compared with the conventional equal-length anti-shake scheme that uses the same window length evenly during the wave period and the rapid rise period, the wave period filters model noise more thoroughly and the rapid rise period shortens the response delay. This dual-speed behavior was often underestimated before the system went online, but after operation, it became a key contribution to supporting the timeliness of high-alert-level video verification.
[0028] Example 2: This embodiment, based on Embodiment 1, provides an end-to-end time series prediction system for early warning of flash floods in small watersheds, such as... Figure 8 As shown, it includes: Multi-source data processing module: used to collect raw measurement values from multi-source heterogeneous sensing devices deployed in a small watershed to construct a multi-source heterogeneous raw data stream with dual timestamps, and sequentially perform spatiotemporal interpolation alignment and missing item completion on the multi-source heterogeneous raw data stream with dual timestamps to generate a complete data matrix; The flood peak forecast and early warning module is used to input the complete data matrix into the end-to-end time series prediction branch for multi-time-scale rolling prediction, output the flood peak forecast result, and perform hierarchical mapping based on the flood peak forecast result to obtain four-level forecast and early warning labels; The PTZ resource scheduling module is used to perform spatial radius coverage query and time dimension anti-shake filtering for the four-level forecast and early warning indicators, respectively to obtain the candidate camera PTZ resource sorting list and the effective early warning event sequence. Based on the candidate camera PTZ resource sorting list and the effective early warning event sequence, the module performs hierarchical preemptive PTZ lock signaling interaction to determine the target camera PTZ. Early warning and handling work order module: Used to establish a concurrent request conflict queue, output conflict arbitration results to control the duty terminal to pull video images from the target camera pan-tilt unit, and bind the video images with the valid early warning event sequence and flood peak forecast results to generate an early warning and handling work order.
[0029] Furthermore, in the multi-source data processing module, the method for constructing a multi-source heterogeneous raw data stream with dual timestamps includes: Each multi-source heterogeneous sensing device is bound to a measurement station location. The original measurement value is simultaneously appended with the device timestamp generated by the local clock of the multi-source heterogeneous sensing device and the receiving timestamp stamped by the data access server, and combined to generate a multi-source heterogeneous raw data stream with dual timestamps.
[0030] The spatiotemporal interpolation alignment method includes: The device timestamp and receiving timestamp of each multi-source heterogeneous sensing device in the multi-source heterogeneous raw data stream with dual timestamps are extracted. The difference between the device timestamp and the receiving timestamp is used as the state variable to construct a first-order Kalman filter. A common sampling grid composed of equally spaced time sequences is set on the time axis as an alignment reference. The first-order Kalman filter is used to perform asynchronous spatiotemporal interpolation alignment processing of the multi-source heterogeneous raw data stream with dual timestamps oriented towards the common sampling grid, and the aligned multi-source heterogeneous data matrix is output.
[0031] The method for performing missing item completion includes: Using the locations of the stations bound to the multi-source heterogeneous sensing devices as graph nodes, the edge weights of the graph are determined based on the geographical distance between the station locations to construct a spatial correlation graph of the multi-source heterogeneous sensing devices. Missing items in the aligned multi-source heterogeneous data matrix are located, and the aligned multi-source heterogeneous data matrix is input into a graph neural network driven by the spatial correlation graph of the multi-source heterogeneous sensing devices. Feature reconstruction is performed on the missing items in the aligned multi-source heterogeneous data matrix to generate a complete data matrix.
[0032] Furthermore, in the flood peak forecasting and early warning module, the method for outputting the flood peak forecasting results includes: The complete data matrix is truncated to the length of the forecast window, and multiple sets of input tensors are constructed. These multiple sets of input tensors are pushed into the corresponding end-to-end time series prediction branches to perform rolling updates. The forecast values returned by each end-to-end time series prediction branch are integrated to output the flood peak forecast result. The flood peak forecast result includes the flood peak water level forecast value, the surface flow velocity forecast value, and the flood peak occurrence time forecast value.
[0033] The method for obtaining four-level forecast and early warning identifiers by performing hierarchical mapping includes: Based on the comparison between the flood peak water level forecast value in the flood peak forecast results and the flood limit water level threshold pre-configured at the station location, a hierarchical mapping is performed to generate corresponding four-level forecast and early warning labels; the flood limit water level threshold includes the early warning water level threshold, the prepared evacuation water level threshold, and the immediate evacuation water level threshold.
[0034] The methods and systems of this application may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the method is for illustrative purposes only, and the steps of the method of this application are not limited to the order specifically described above, unless otherwise specifically stated.
[0035] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.
[0036] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements 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 small watershed flash flood early warning method based on end-to-end time series prediction, characterized in that, The method includes: The raw measurement values of multi-source heterogeneous sensing devices deployed in a small watershed are collected to construct a multi-source heterogeneous raw data stream with double timestamps. The multi-source heterogeneous raw data stream with double timestamps is sequentially subjected to spatiotemporal interpolation alignment and missing term completion to generate a complete data matrix. The complete data matrix is input into the end-to-end time series prediction branch for multi-timescale rolling prediction, and the flood peak forecast result is output. Based on the flood peak forecast result, a hierarchical mapping is performed to obtain a four-level forecast and warning label. The method for sequentially performing spatiotemporal interpolation alignment on a multi-source heterogeneous raw data stream with dual timestamps includes: extracting the device timestamp and receiving timestamp of each multi-source heterogeneous sensing device in the multi-source heterogeneous raw data stream with dual timestamps; using the difference between the device timestamp and the receiving timestamp as a state variable to construct a first-order Kalman filter; setting a common sampling grid composed of equally spaced time sequences on the time axis as an alignment reference; using the first-order Kalman filter to perform asynchronous spatiotemporal data interpolation alignment processing on the multi-source heterogeneous raw data stream with dual timestamps oriented towards the common sampling grid; and outputting the aligned multi-source heterogeneous data matrix. For the Level 4 forecast and early warning indicators, spatial radius coverage query and temporal anti-shake filtering are performed to obtain a candidate camera PTZ resource ranking list and a valid early warning event sequence. Based on the candidate camera PTZ resource ranking list and the valid early warning event sequence, hierarchical preemptive PTZ lock signaling interaction is performed to determine the target camera PTZ. A concurrent request conflict queue is established to output the conflict arbitration result to control the duty terminal to pull the video image from the target camera PTZ. The video image is bound with the valid early warning event sequence and the flood peak forecast result to generate an early warning handling work order.
2. The small watershed flash flood early warning method based on end-to-end time series prediction according to claim 1, characterized in that, The method for constructing a multi-source heterogeneous raw data stream with dual timestamps includes: Each multi-source heterogeneous sensing device is bound to a measurement station location. The original measurement value is simultaneously appended with the device timestamp generated by the local clock of the multi-source heterogeneous sensing device and the receiving timestamp stamped by the data access server, and combined to generate a multi-source heterogeneous raw data stream with dual timestamps.
3. The small watershed flash flood early warning method based on end-to-end time series prediction according to claim 2, characterized in that, The method for performing missing item completion includes: Using the locations of the stations bound to the multi-source heterogeneous sensing devices as graph nodes, the edge weights of the graph are determined based on the geographical distance between the station locations to construct a spatial correlation graph of the multi-source heterogeneous sensing devices. Missing items in the aligned multi-source heterogeneous data matrix are located, and the aligned multi-source heterogeneous data matrix is input into a graph neural network driven by the spatial correlation graph of the multi-source heterogeneous sensing devices. Feature reconstruction is performed on the missing items in the aligned multi-source heterogeneous data matrix to generate a complete data matrix.
4. The method for early warning of flash floods in small watersheds based on end-to-end time series prediction according to claim 1, characterized in that, The method for outputting flood peak forecast results includes: The complete data matrix is truncated to the length of the forecast window, and multiple sets of input tensors are constructed. These multiple sets of input tensors are pushed into the corresponding end-to-end time series prediction branches to perform rolling updates. The forecast values returned by each end-to-end time series prediction branch are integrated to output the flood peak forecast result. The flood peak forecast result includes the flood peak water level forecast value, the surface flow velocity forecast value, and the flood peak occurrence time forecast value.
5. The method for early warning of flash floods in small watersheds based on end-to-end time series prediction according to claim 4, characterized in that, The method for obtaining four-level forecast and early warning identifiers by performing hierarchical mapping includes: Based on the comparison between the flood peak water level forecast value in the flood peak forecast results and the flood limit water level threshold pre-configured at the station location, a hierarchical mapping is performed to generate corresponding four-level forecast and early warning labels; the flood limit water level threshold includes the early warning water level threshold, the prepared evacuation water level threshold, and the immediate evacuation water level threshold.
6. The method for early warning of flash floods in small watersheds based on end-to-end time series prediction according to claim 5, characterized in that, The determination logic for the graded mapping is as follows: when the flood peak water level forecast value in the flood peak forecast result is greater than or equal to the immediate evacuation water level threshold, it is mapped to a red warning; when the flood peak water level forecast value is greater than or equal to the prepared evacuation water level threshold and less than the immediate evacuation water level threshold, it is mapped to an orange warning; when the flood peak water level forecast value is greater than or equal to the warning water level threshold and less than the prepared evacuation water level threshold, it is mapped to a yellow warning; and when the flood peak water level forecast value is less than the warning water level threshold, it is mapped to a blue warning.
7. The method for early warning of flash floods in small watersheds based on end-to-end time series prediction according to claim 5, characterized in that, The method for spatial radius coverage query includes: For each Level 4 forecast and early warning sign associated with a monitoring station location, a radius coverage query is performed to identify candidate camera PTZ resources. The candidate camera PTZ resources are then sorted in ascending order of their geographical distance from the monitoring station location, generating a sorted list of candidate camera PTZ resources.
8. The method for early warning of flash floods in small watersheds based on end-to-end time series prediction according to claim 5, characterized in that, The time-dimensional anti-shake filtering method includes: Within a preset anti-shake time window, the Level 4 forecast and warning indicators are counted by sliding. Level 4 forecast and warning indicators that have not reached the effective counting threshold are filtered out, and the Level 4 forecast and warning indicators that have reached the effective counting threshold are summarized into a valid warning event sequence.
9. A method for early warning of flash floods in small watersheds based on end-to-end time series prediction according to claim 8, characterized in that, When the highest warning level of a Level 4 forecast warning sign appears within the preset anti-shake time window is higher than the warning level of the most recently added valid warning event in the valid warning event sequence, the immediate release rule is triggered. The immediate release rule skips the waiting time for the accumulated sliding count and directly adds the Level 4 forecast warning sign with the highest warning level as a valid warning event to the valid warning event sequence.
10. A method for early warning of flash floods in small watersheds based on end-to-end time series prediction according to claim 7, characterized in that, The method for determining the target camera gimbal through hierarchical preemptive gimbal lock signaling interaction includes: Maintain a hierarchical preemptive PTZ lock for each camera PTZ. Based on the warning level carried by the effective warning event sequence, initiate a locking request to each camera PTZ along the candidate camera PTZ resource sorting list, perform hierarchical preemptive PTZ lock signaling interaction, determine the PTZ lock occupancy status of each camera PTZ, and authorize the locking request to lock the target camera PTZ.
11. The method for early warning of flash floods in small watersheds based on end-to-end time series prediction according to claim 10, characterized in that, The method for establishing a concurrent request conflict queue and outputting the conflict arbitration result includes: Multiple locking requests arriving at the target camera PTZ at the same time are collected to establish a concurrent request conflict queue. Based on the PTZ lock occupancy status, multiple locking requests in the concurrent request conflict queue are sorted by multiple dimensions and the lock handover is managed. The conflict arbitration result for multiple locking requests is output.
12. A small watershed flash flood early warning system based on end-to-end time series prediction, used to implement the small watershed flash flood early warning method based on end-to-end time series prediction as described in any one of claims 1-11, characterized in that, The system includes: Multi-source data processing module: used to collect raw measurement values from multi-source heterogeneous sensing devices deployed in a small watershed to construct a multi-source heterogeneous raw data stream with dual timestamps, and sequentially perform spatiotemporal interpolation alignment and missing item completion on the multi-source heterogeneous raw data stream with dual timestamps to generate a complete data matrix; The flood peak forecast and early warning module is used to input the complete data matrix into the end-to-end time series prediction branch for multi-time-scale rolling prediction, output the flood peak forecast result, and perform hierarchical mapping based on the flood peak forecast result to obtain four-level forecast and early warning labels; The PTZ resource scheduling module is used to perform spatial radius coverage query and time dimension anti-shake filtering for the four-level forecast and early warning indicators, respectively to obtain the candidate camera PTZ resource sorting list and the effective early warning event sequence. Based on the candidate camera PTZ resource sorting list and the effective early warning event sequence, the module performs hierarchical preemptive PTZ lock signaling interaction to determine the target camera PTZ. Early warning and handling work order module: Used to establish a concurrent request conflict queue, output conflict arbitration results to control the duty terminal to pull video images from the target camera pan-tilt unit, and bind the video images with the valid early warning event sequence and flood peak forecast results to generate an early warning and handling work order.
Citation Information
Patent Citations
Small watershed flash flood forecasting and warning method and device, electronic equipment, and storage medium
CN118397797B
Mountain torrent disaster early warning system based on deep learning and air-space-ground integrated perception
CN121963412A
Timestamp-free synchronous frequency offset tracking method based on dynamic response
CN110572232A
Water area detection monitoring system
CN110611791A
