Rainfall intensity rapid identification method and system based on multi-sensor data fusion

Abstract

The invention discloses a rainfall intensity rapid identification method and system based on multi-sensor data fusion, and relates to the technical field of hydro-meteorological monitoring and urban drainage scheduling, and the method comprises the steps: obtaining a historical rainfall data set, drawing a peak curve, and extracting an evolution sequence feature before a peak value; correlation analysis is carried out on the data and complete rainfall process characteristics to establish a historical rule model; rainfall intensity data are collected in real time through multiple sensors, and space-time dimension features are fused; matching the fusion features with historical features, and screening target historical events; constructing a rainfall reproduction prediction model to reconstruct a current rainfall evolution process; and finally, calculating an inflow load based on a reconstruction result, and generating a dynamic drainage scheduling scheme. According to the invention, accurate prediction of the rainfall process and intelligent scheduling of the drainage system are realized, and the urban waterlogging prevention capability is improved.

Description

Technical Field

[0001] This invention relates to the field of hydrological and meteorological monitoring and urban drainage scheduling technology, specifically to a method and system for rapid identification of rainfall intensity based on multi-sensor data fusion. Background Technology

[0002] Frequent urban flooding severely impacts urban infrastructure and people's lives. The effective operation of urban drainage systems is crucial for preventing flooding. However, urban drainage systems typically employ fixed scheduling schemes, failing to dynamically adjust according to actual rainfall conditions, resulting in low system efficiency.

[0003] Traditional rainfall monitoring methods mainly rely on single-point rain gauges or weather radars, which suffer from low spatial resolution and data acquisition delays. Existing rainfall forecasting methods often only consider time series characteristics, neglecting spatial distribution information, and have high computational complexity, making it difficult to meet the needs of real-time forecasting.

[0004] Current technologies for analyzing rainfall processes are largely limited to statistical methods, lacking a deep understanding of the evolutionary patterns of rainfall events. In identifying rainfall peaks, they primarily rely on fixed thresholds or simple mathematical models, failing to accurately grasp the dynamic changes in rainfall intensity and impacting forecast accuracy and warning timeliness.

[0005] Current urban drainage dispatching systems generally suffer from problems such as delayed response and insufficient foresight. Due to the lack of accurate prediction of rainfall events, it is difficult to adjust drainage strategies in a timely manner, resulting in poor system dispatching effectiveness and affecting the city's flood control capabilities. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for rapid identification of rainfall intensity based on multi-sensor data fusion, aiming to solve at least one of the technical problems existing in the prior art.

[0007] The technical solution of this invention is: a method for rapid identification of rainfall intensity based on multi-sensor data fusion, comprising the following steps: Historical precipitation datasets for the study area are obtained, peak precipitation intensity curves are plotted, and precipitation events are segmented based on the peak precipitation curves to obtain pre-peak evolution sequence characteristics. A correlation analysis was conducted between the pre-peak evolution sequence characteristics and the characteristics of the complete rainfall process to establish a historical rule correspondence model. The characteristics of the complete rainfall process include pre-peak evolution characteristics, peak characteristics, and post-peak attenuation characteristics. The system collects rainfall intensity data of the current rainfall event in real time by multiple rainfall sensors, generates the current rainfall evolution sequence, and extracts the rainfall intensity change features in the time dimension and the sensor distribution features in the spatial dimension to form a fused feature. The fused features are matched with the pre-peak evolution sequence features, and the target set of historical rainfall events is selected according to the historical rule corresponding model and the corresponding complete rainfall process features are extracted. A rainfall recurrence prediction model is constructed based on the characteristics of the complete rainfall process, and the rainfall intensity evolution process of the current rainfall event is reconstructed through the rainfall recurrence prediction model. Based on the reconstructed rainfall intensity evolution process, the inflow load change of the urban drainage system is calculated, and a dynamic drainage scheduling scheme is generated.

[0008] Historical precipitation datasets for the study area were obtained, and peak precipitation intensity curves were plotted. Based on the peak precipitation curves, precipitation events were segmented to obtain pre-peak evolution sequence characteristics, including: Perform time series analysis on historical precipitation datasets, construct an adaptive sliding window, calculate the mean and variance of precipitation intensity data within the adaptive sliding window, generate a peak identification threshold, and use the peak identification threshold to identify local maxima of precipitation intensity within the adaptive sliding window. Calculate the time interval and intensity variation between local maxima of rainfall intensity to construct a state variable matrix; The significance index is calculated based on the state quantity matrix, and the local maxima of rainfall intensity are sorted. The maxima with significance index higher than the mean are selected as the rainfall peak points, and the rainfall peak points are connected to form the rainfall intensity peak curve. The fluctuation trend of the peak rainfall intensity curve is analyzed, and the fluctuation inflection point position is extracted from the peak rainfall intensity curve. The fluctuation inflection point position is used as the segment boundary point of the rainfall event, and the pre-peak evolution sequence features of the rainfall event are extracted based on the segment boundary point.

[0009] The correlation analysis between the pre-peak evolution sequence characteristics and the characteristics of the complete rainfall process was used to establish a historical rule correspondence model, including: Extract the rainfall feature vector from the pre-peak evolution sequence features, retrieve the corresponding time period feature vector of the rainfall feature vector in the complete rainfall process features, and calculate the similarity value between the rainfall feature vector and the corresponding time period feature vector; Historical rainfall events are grouped based on similarity values, the state transition probability of rainfall events within each group is calculated, and the state transition matrix of rainfall events is constructed using the state transition probability. Extract the state transition patterns from the state transition matrix of rainfall events and construct a set of state transition rules; The rules in the state transition rule set are subjected to reliability assessment, a rule assessment score is generated, transition rules with assessment scores higher than a preset threshold are selected, and a historical rule correspondence model is established.

[0010] Rainfall intensity data for the current rainfall event is collected in real time by multiple rainfall sensors, generating the current rainfall evolution sequence. The temporal dimensions of rainfall intensity variation and the spatial dimensions of sensor distribution characteristics are extracted to form fused features, including: A timestamp calibration operation is performed on the rainfall intensity data collected by the rainfall sensor to generate calibrated rainfall intensity data. Based on the calibrated rainfall intensity data, a current rainfall evolution sequence is generated, and the time dimension rainfall intensity change features are extracted from the current rainfall evolution sequence. The spatial coordinates of the rainfall sensor are obtained, and the calibrated rainfall intensity data is combined with the spatial coordinates to construct a spatial distribution matrix. The sensor distribution features in the spatial dimension are extracted from the spatial distribution matrix. The time-dimensional rainfall intensity variation characteristics and the spatial sensor distribution characteristics are fused to generate fused features.

[0011] The fused features are matched with the pre-peak evolution sequence features. Based on the historical rules corresponding to the model, the target set of historical rainfall events is selected and the corresponding complete rainfall process features are extracted, including: The fused features and the pre-peak evolution sequence features are used to perform feature mapping calculations to generate a feature similarity matrix, and a rainfall feature matching vector is constructed based on the feature similarity matrix. Based on the mapping rules in the historical rule correspondence model, the similarity of the rainfall feature matching vector is evaluated, the matching degree between each historical rainfall event and the current rainfall event is calculated, and a historical rainfall event matching score is generated. A filtering threshold is set based on the matching score of historical rainfall events, and historical rainfall events with matching scores higher than the filtering threshold are selected to form a target set of historical rainfall events; Based on the order of historical rainfall events in the target historical rainfall event set, a retrieval sequence is constructed by combining the matching score, and the corresponding complete rainfall process features are extracted from the historical rainfall process feature data according to the retrieval sequence.

[0012] A rainfall recurrence prediction model is constructed based on the characteristics of a complete rainfall process. The rainfall intensity evolution process of the current rainfall event is reconstructed through the rainfall recurrence prediction model, including: The complete rainfall process features are decomposed according to the temporal structure to generate a temporal feature mapping vector; Construct a prediction model structure that includes a temporal coding unit, a feature association unit, and a prediction decoding unit; The temporal feature mapping vector is input into the temporal coding unit to extract temporal correlation features; Temporal correlation features are input into the feature association unit to establish a feature mapping relationship, and a prediction parameter matrix is ​​generated based on the feature mapping relationship. The prediction parameter matrix is ​​input into the prediction decoding unit to construct a rainfall recurrence prediction model; The feature sequence of the current rainfall event is input into the rainfall recurrence prediction model. The time-series coding unit generates coded features. The coded features are input into the feature association unit to calculate the feature mapping value. The feature mapping value is input into the prediction decoding unit to generate the rainfall intensity change sequence. The rainfall intensity change sequence is discretized according to the time step to generate a discrete time sequence. The rainfall intensity value corresponding to the discrete time sequence is calculated. The rainfall intensity value is recombined according to the temporal relationship to reconstruct the rainfall intensity evolution process of the current rainfall event.

[0013] Based on the reconstructed rainfall intensity evolution, the inflow load changes of the urban drainage system are calculated, and a dynamic drainage scheduling scheme is generated, including: The reconstructed complete rainfall intensity evolution process is converted into a time series. The runoff coefficient is calculated based on the land cover data to generate inflow load changes. The drainage system load curve is constructed based on the inflow load changes. Hydraulic calculations are performed on the load curve of the drainage system to obtain the hydraulic parameters of the pipe network nodes. Based on the hydraulic parameters, a pipe network hydraulic state matrix is ​​constructed to generate a drainage capacity assessment value. Based on the drainage capacity assessment value and the pipeline status matrix, the pump station start-up and shutdown parameters and the gate opening parameters are calculated, and the pump station start-up and shutdown parameters and the gate opening parameters are combined to generate the drainage facility operation parameters. The operating parameters of the drainage facilities are divided into multiple time periods according to fixed time intervals. Pump station operation instructions and gate adjustment instructions are generated for each time period to form a time-segmented scheduling instruction sequence. A dynamic drainage scheduling scheme is constructed based on the time-segmented scheduling instruction sequence.

[0014] This invention provides a rapid rainfall intensity identification system based on multi-sensor data fusion, the system comprising: The historical data processing module is used to acquire historical precipitation datasets for the study area, plot peak curves of rainfall intensity, segment rainfall events based on peak curves, and obtain pre-peak evolution sequence characteristics. The rule modeling module is used to perform correlation analysis between the pre-peak evolution sequence features and the complete rainfall process features to establish a historical rule correspondence model. The complete rainfall process features include pre-peak evolution features, peak features, and post-peak decay features. The real-time monitoring module includes multiple rainfall sensors, which are used to collect rainfall intensity data of the current rainfall event in real time, generate the current rainfall evolution sequence, and extract the rainfall intensity change features in the time dimension and the sensor distribution features in the spatial dimension to form fused features; The feature matching module is used to match the fused features with the pre-peak evolution sequence features, filter out the target set of historical rainfall events according to the historical rule corresponding model, and extract the corresponding complete rainfall process features. The prediction modeling module is used to construct a rainfall recurrence prediction model based on the characteristics of the complete rainfall process, and to reconstruct the rainfall intensity evolution process of the current rainfall event through the rainfall recurrence prediction model; The scheduling scheme generation module is used to calculate the inflow load changes of the urban drainage system based on the reconstructed rainfall intensity evolution process and generate a dynamic drainage scheduling scheme.

[0015] One technical solution provided in this embodiment of the invention is an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps in any of the aforementioned methods.

[0016] One technical solution provided in this embodiment of the invention is a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the steps in any of the aforementioned methods.

[0017] This invention establishes a rainfall evolution model by fusing real-time monitoring data from multiple sensors with historical rainfall patterns. This model accurately identifies rainfall intensity variation characteristics, improving the spatiotemporal resolution and accuracy of rainfall prediction. The method utilizes matching analysis between pre-peak rainfall evolution sequence features and historical data to significantly enhance the real-time performance and accuracy of the prediction model. By establishing the correspondence between complete rainfall process characteristics and pre-peak evolution features, rapid reconstruction of the rainfall process is achieved, providing a reliable basis for drainage system scheduling. Based on the reconstructed rainfall evolution process, changes in inflow load can be predicted in a timely manner, and drainage schemes can be dynamically optimized, effectively improving the operational efficiency and flood control capacity of urban drainage systems and reducing the risk of urban flooding. Attached Figure Description

[0018] Figure 1 A flowchart of a method for rapid identification of rainfall intensity based on multi-sensor data fusion provided in an embodiment of the present invention; Figure 2 The flowchart for generating the dynamic drainage scheduling scheme in this embodiment of the invention is shown. Detailed Implementation

[0019] like Figure 1 As shown, Figure 1 This is a flowchart of a method for rapid identification of rainfall intensity based on multi-sensor data fusion provided in an embodiment of the present invention. The method includes the following steps: Historical precipitation datasets for the study area are obtained, peak precipitation intensity curves are plotted, and precipitation events are segmented based on the peak precipitation curves to obtain pre-peak evolution sequence characteristics. A correlation analysis was conducted between the pre-peak evolution sequence characteristics and the characteristics of the complete rainfall process to establish a historical rule correspondence model. The characteristics of the complete rainfall process include pre-peak evolution characteristics, peak characteristics, and post-peak attenuation characteristics. The system collects rainfall intensity data of the current rainfall event in real time by multiple rainfall sensors, generates the current rainfall evolution sequence, and extracts the rainfall intensity change features in the time dimension and the sensor distribution features in the spatial dimension to form a fused feature. The fused features are matched with the pre-peak evolution sequence features, and the target set of historical rainfall events is selected according to the historical rule corresponding model and the corresponding complete rainfall process features are extracted. A rainfall recurrence prediction model is constructed based on the characteristics of the complete rainfall process, and the rainfall intensity evolution process of the current rainfall event is reconstructed through the rainfall recurrence prediction model. Based on the reconstructed rainfall intensity evolution process, the inflow load change of the urban drainage system is calculated, and a dynamic drainage scheduling scheme is generated.

[0020] Historical precipitation datasets for the study area were obtained, and peak precipitation intensity curves were plotted. Based on the peak precipitation curves, precipitation events were segmented to obtain pre-peak evolution sequence characteristics, including: Perform time series analysis on historical precipitation datasets, construct an adaptive sliding window, calculate the mean and variance of precipitation intensity data within the adaptive sliding window, generate a peak identification threshold, and use the peak identification threshold to identify local maxima of precipitation intensity within the adaptive sliding window. Calculate the time interval and intensity variation between local maxima of rainfall intensity to construct a state variable matrix; The significance index is calculated based on the state quantity matrix, and the local maxima of rainfall intensity are sorted. The maxima with significance index higher than the mean are selected as the rainfall peak points, and the rainfall peak points are connected to form the rainfall intensity peak curve. The fluctuation trend of the peak rainfall intensity curve is analyzed, and the fluctuation inflection point position is extracted from the peak rainfall intensity curve. The fluctuation inflection point position is used as the segment boundary point of the rainfall event, and the pre-peak evolution sequence features of the rainfall event are extracted based on the segment boundary point.

[0021] A time-series analysis was performed on historical precipitation datasets, constructing an adaptive sliding window. The size of the adaptive sliding window was determined based on the temporal resolution of the precipitation data and the typical duration of precipitation events. The window size was not fixed but automatically adjusted according to the rate of change in precipitation intensity. When precipitation intensity changed drastically, the window became smaller to capture rapid changes; when precipitation intensity was relatively stable, the window became larger to reduce the influence of noise. The adaptive window size typically ranged from 1 to 24 hours, depending on the precipitation characteristics of the study area.

[0022] Within an adaptive sliding window, the mean and variance of rainfall intensity data are calculated to generate a peak identification threshold. The mean reflects the average level of rainfall intensity within the window, while the variance reflects the degree of fluctuation in rainfall intensity. The peak identification threshold is calculated by multiplying the mean and variance, with the product coefficient determined based on the rainfall characteristics of the study area, typically between 1.5 and 3. A higher product coefficient is used when rainfall intensity varies significantly in the study area, and a lower product coefficient is used when rainfall intensity varies less. Using the generated peak identification threshold, local maxima of rainfall intensity are identified within the adaptive sliding window. A local maximum is a point where, within a certain time range, the rainfall intensity is higher than that of surrounding time points and also exceeds the threshold.

[0023] The time interval and intensity variation between local maxima of rainfall intensity are calculated to construct a state variable matrix. The time interval refers to the time difference between two adjacent local maxima, reflecting the frequency of intensity peaks during rainfall. The intensity variation refers to the difference in rainfall intensity between two adjacent local maxima, reflecting the magnitude of the intensity change. Each row of the state variable matrix represents a local maximum, containing information such as the time, intensity, time interval from the previous point, and intensity variation at that point.

[0024] A significance index is calculated based on the state variable matrix, and the local maxima of rainfall intensity are ranked. The significance index comprehensively considers the intensity value, time interval value, and intensity change value of the local maxima. During the index calculation, these three parameters are normalized and then weighted and summed. The intensity value has a weight of 0.5, the time interval value has a weight of 0.3, and the intensity change value has a weight of 0.2. Maximum values ​​of the significance index above the mean are selected as rainfall peak points, and these peak points are connected to form a rainfall intensity peak curve.

[0025] This study analyzes the fluctuation trend of the peak rainfall intensity curve and extracts the inflection points from it. An inflection point is a point where the sign of the slope of the peak curve changes, indicating a turning point where rainfall intensity changes from increasing to decreasing or vice versa. The inflection point identification uses the sliding window difference method to calculate the slope change between adjacent peak points. These inflection points are then used as segmentation boundary points for rainfall events, and the pre-peak evolution sequence features of the rainfall events are extracted based on these segmentation boundary points.

[0026] Taking a specific study area as an example, precipitation data fused from meteorological stations and radar was analyzed. The annual precipitation data for this area was processed and analyzed using an initial adaptive window of 6 hours. Within the window, the mean rainfall intensity was 5 mm / h, and the variance was 4 mm / h. A peak intensity threshold of 15 mm / h was calculated using a product factor of 2.5. Multiple local maxima were identified using this threshold, and the time intervals and intensity changes between these points were calculated to construct a state matrix. After calculating significance indicators, 25 rainfall peak points were identified and connected to form a rainfall intensity peak curve. The fluctuation trend of the peak curve was analyzed, and 18 fluctuation inflection points were extracted as segmentation boundary points for rainfall events, ultimately identifying 12 complete rainfall events. For each rainfall event, evolutionary sequence features such as rainfall intensity, duration, and growth rate before the peak were extracted to provide a basis for subsequent rainfall forecasting and early warning.

[0027] This invention utilizes an adaptive sliding window technique to automatically adjust the analysis window size based on changes in rainfall intensity, adapting to the characteristics of different types of rainfall events. By leveraging multi-sensor fusion data, it improves the accuracy and spatiotemporal coverage of rainfall intensity estimation. Through the identification of rainfall intensity peak curves and fluctuation inflection points, it achieves precise segmentation of rainfall events, avoiding the limitations of traditional fixed threshold methods. The extraction of pre-peak evolutionary sequence features of rainfall events provides a scientific basis for rainfall early warning and disaster prevention and mitigation, significantly improving the predictive ability and early warning time of extreme rainfall events, and has important application value for water resource management and flood control and disaster reduction.

[0028] The correlation analysis between the pre-peak evolution sequence characteristics and the characteristics of the complete rainfall process was used to establish a historical rule correspondence model, including: Extract the rainfall feature vector from the pre-peak evolution sequence features, retrieve the corresponding time period feature vector of the rainfall feature vector in the complete rainfall process features, and calculate the similarity value between the rainfall feature vector and the corresponding time period feature vector; Historical rainfall events are grouped based on similarity values, the state transition probability of rainfall events within each group is calculated, and the state transition matrix of rainfall events is constructed using the state transition probability. Extract the state transition patterns from the state transition matrix of rainfall events and construct a set of state transition rules; The rules in the state transition rule set are subjected to reliability assessment, a rule assessment score is generated, transition rules with assessment scores higher than a preset threshold are selected, and a historical rule correspondence model is established.

[0029] Precipitation feature vectors are extracted from the pre-peak evolution sequence features. These feature vectors contain multi-dimensional information, primarily including the rate of change of rainfall intensity, cumulative rainfall, duration, and frequency of intensity fluctuations. Pre-peak evolution sequence features refer to the evolution of rainfall patterns over a period before the peak rainfall intensity. Feature extraction employs a sliding time window method, calculating features from rainfall data at different time points before the peak to form a time-series feature vector. The window size is adjusted according to the timescale of the rainfall event, typically 1 to 6 hours before the peak. Different feature weights are used for fusion of different types of sensor data: meteorological station data has a weight of 0.4, radar data 0.35, and satellite data 0.25. This multi-sensor data fusion approach compensates for the limitations of a single data source and improves the accuracy of the feature vectors.

[0030] The process retrieves the corresponding time-period feature vectors from the complete rainfall event features. Complete rainfall event features refer to the feature description of the entire event from the start to the end of rainfall, including features in the pre-peak, peak, and post-peak stages. The corresponding time-period feature vector refers to the historical rainfall event feature vector that corresponds to the currently extracted pre-peak evolution sequence feature vector at its temporal location. The retrieval process uses a historical rainfall event database, which contains rainfall event records and feature vectors from the past several years in the study area. A sliding comparison method is used, comparing the current pre-peak evolution sequence feature vector with feature vectors in the historical database one by one. When calculating the similarity value between the rainfall feature vector and the corresponding time-period feature vector, a cosine similarity method is used, which considers the weight differences of each dimension of the feature vector. The similarity value ranges from 0 to 1, with values ​​closer to 1 indicating greater similarity between the two feature vectors.

[0031] Historical rainfall events are grouped based on similarity values ​​using a hierarchical clustering algorithm. Each rainfall event is treated as an independent category, and the most similar categories are gradually merged based on similarity values ​​until a preset number of categories is reached. The number of groups is determined by silhouette coefficient evaluation, generally ranging from 3 to 7 groups, depending on the diversity of rainfall types in the study area. Taking a certain study area as an example, by analyzing 500 historical rainfall events, five categories were ultimately identified: short-duration heavy rainfall, long-duration moderate rainfall, persistent weak rainfall, multi-peak rainfall, and rapidly intensifying rainfall. For rainfall events within each group, state transition probabilities are calculated. A state is defined as a different interval of rainfall intensity, such as light rain, moderate rain, heavy rain, and torrential rain. The state transition probability refers to the probability of rainfall changing from one state to another, calculated by statistically analyzing the frequency of state transitions in historical rainfall events within the group. A state transition matrix of rainfall events is constructed using the state transition probabilities.

[0032] State transition patterns are extracted from the state transition matrix of rainfall events to construct a set of state transition rules. State transition patterns refer to fixed patterns of state transitions during rainfall, such as the transition from light rain to moderate rain, and then to heavy rain. The extraction process employs a frequent pattern mining algorithm, which can discover recurring patterns in the state transition sequence. The algorithm sets a minimum support of 0.15, meaning that a transition pattern is considered valid only if it occurs at least 15% of the time within a given group. The constructed set of state transition rules contains the state transition patterns found in each group, and each rule includes three parts: preconditions, transition sequence, and probability of occurrence.

[0033] The reliability of rules in the state transition rule set was evaluated using cross-validation. Historical data was divided into training and validation sets in a 7:3 ratio. Rules were constructed using the training set, and their accuracy was tested on the validation set. Reliability evaluation metrics included rule accuracy, coverage, and stability. Accuracy refers to the proportion of correctly predicted events, coverage refers to the proportion of rainfall events to which the rule can be applied, and stability refers to the consistency of the rule's performance across different time periods. The generated rule evaluation score was a weighted average of these three metrics, with weights of 0.5, 0.3, and 0.2, respectively. Taking a specific rainfall type in the study area as an example, a transition rule was extracted: "When light rain lasts for 2 hours and the rainfall intensity steadily increases, there is an 80% probability that it will turn into moderate rain in the next hour." The evaluation of this rule showed an accuracy of 0.82, a coverage of 0.65, a stability of 0.78, and a comprehensive evaluation score of 0.77. Transition rules with evaluation scores higher than a preset threshold (typically set at 0.7) were selected; rules above this threshold were considered to have sufficient reliability. The selected high-reliability rules are used to build a historical rule correspondence model. This model can match historical rules with the current rainfall pre-peak evolution sequence characteristics to predict the subsequent development trend of rainfall.

[0034] This invention establishes a highly reliable historical rule correspondence model by correlating the pre-peak evolution sequence features with the features of the complete rainfall process, demonstrating significant technical advantages. Based on multi-sensor fusion data, it improves the accuracy and comprehensiveness of rainfall feature extraction, overcoming the limitations of single-sensor data in terms of coverage and precision. Through similarity calculation and group clustering, it achieves accurate identification and classification of different types of rainfall events, improving the targeting of rainfall pattern mining. The construction of state transition rules and the reliability assessment mechanism ensure the credibility of the model's prediction results.

[0035] Rainfall intensity data for the current rainfall event is collected in real time by multiple rainfall sensors, generating the current rainfall evolution sequence. The temporal dimensions of rainfall intensity variation and the spatial dimensions of sensor distribution characteristics are extracted to form fused features, including: A timestamp calibration operation is performed on the rainfall intensity data collected by the rainfall sensor to generate calibrated rainfall intensity data. Based on the calibrated rainfall intensity data, a current rainfall evolution sequence is generated, and the time dimension rainfall intensity change features are extracted from the current rainfall evolution sequence. The spatial coordinates of the rainfall sensor are obtained, and the calibrated rainfall intensity data is combined with the spatial coordinates to construct a spatial distribution matrix. The sensor distribution features in the spatial dimension are extracted from the spatial distribution matrix. The time-dimensional rainfall intensity variation characteristics and the spatial sensor distribution characteristics are fused to generate fused features.

[0036] Rainfall sensors include various types of sensing devices such as ground rain gauges, weather radars, and satellite remote sensing equipment. These sensors collect rainfall intensity data with different temporal resolutions and sampling frequencies. Ground rain gauges typically collect data every 5 or 10 minutes, weather radars complete a scan every 6 minutes, while satellite remote sensing data may be updated every 30 minutes or longer. Due to the inconsistent sampling times, timestamp calibration is required for the rainfall intensity data collected by the rain sensors. Timestamp calibration uses time alignment technology to unify the data from different sensors onto the same time axis. During calibration, for sensor data with high sampling frequencies, the corresponding calibration point is calculated by averaging within a time window; for sensor data with low sampling frequencies, data points at intermediate times are generated using interpolation methods. Taking ground rain gauge and radar data as an example, if the ground rain gauge collects data every 5 minutes and the radar scans every 6 minutes, both can be unified to one data point every 10 minutes, and the calibration data point can be obtained by averaging or interpolating the data within the time window.

[0037] After generating calibrated rainfall intensity data, a current rainfall evolution sequence is generated based on this data. The current rainfall evolution sequence refers to the time series of rainfall intensity from the start of rainfall to the current moment, recording the complete process of rainfall intensity change over time. Data from the first 30 minutes to 2 hours of the current rainfall event can be used as the evolution sequence, with the specific time span determined according to the rainfall type and research objectives. For example, for short-duration heavy rainfall, the first 30 minutes of data can be selected; for continuous rainfall, the first 2 hours of data can be selected. The time-dimensional rainfall intensity change characteristics are extracted from the current rainfall evolution sequence, including the average rainfall intensity, maximum rainfall intensity, rate of change, coefficient of variation, and autocorrelation coefficient. The rate of change of rainfall intensity reflects the speed at which rainfall intensifies or weakens, the coefficient of variation represents the degree of fluctuation in rainfall intensity, and the autocorrelation coefficient reflects the temporal correlation of rainfall intensity. During the extraction process, a sliding time window method is used, with a window size of 10 minutes, sliding for 5 minutes each time, calculating the characteristic parameters within the window to form a time feature sequence.

[0038] The spatial coordinates of rainfall sensors, including longitude, latitude, and elevation information, are obtained through a positioning system. In actual deployment, ground rain gauges are typically distributed at a certain spatial density, with one station per 10 square kilometers in dense areas and one station per 100 square kilometers in sparse areas. Weather radar has a larger coverage area, generally monitoring rainfall within a radius of 200 kilometers. The calibrated rainfall intensity data is combined with the spatial coordinates to construct a spatial distribution matrix. Each row of the spatial distribution matrix represents a sensor, and the columns include the sensor's location coordinates and the corresponding rainfall intensity value. The combination operation considers the characteristics of different types of sensor data, assigning higher weights to ground rain gauge data and adjusting the weights of radar and satellite data based on distance and quality. For data from multiple sensors at the same location, a weighted average method is used for fusion, with weights determined based on data quality and reliability.

[0039] Spatial sensor distribution features, including spatial gradient, clustering, and directionality of rainfall intensity, are extracted from the spatial distribution matrix. The spatial gradient reflects the rate of change of rainfall intensity in space, calculated as the ratio of the difference in rainfall intensity between adjacent sensor locations to their distance. Clustering indicates the spatial concentration of rainfall intensity, calculated using spatial autocorrelation analysis. Directionality reflects the movement direction of rain clouds, determined by analyzing the trend of rainfall intensity changes in different directions. During the extraction process, spatial interpolation methods are used to convert discrete sensor data into a continuous spatial distribution field of rainfall intensity. Commonly used spatial interpolation methods include inverse distance weighting and Kriging interpolation. Taking a specific rainfall event as an example, a spatial distribution matrix is ​​constructed using data from 20 ground rain gauges and one weather radar. The calculated spatial gradient of rainfall intensity is 2 mm / h per kilometer, the clustering index is 0.75, and the main rainfall direction is from northeast to southwest.

[0040] The time-dimensional rainfall intensity variation characteristics and the spatial sensor distribution characteristics are fused to generate a fused feature. The fusion operation employs a feature-level fusion method, merging the feature vectors of the two dimensions into a single comprehensive feature vector. During the fusion process, the relative importance and complementarity of the two types of features are considered, assigning different weights to each feature component. For rapidly changing rainfall events, the time dimension features have higher weights; for spatially unevenly distributed rainfall events, the spatial dimension features have higher weights. In practice, an adaptive weighting method is used to dynamically adjust the feature weights according to the type of rainfall event. The fused feature contains complete information about the rainfall event in both time and space dimensions, comprehensively reflecting the development trend and distribution characteristics of rainfall. Taking a specific rainfall event as an example, the extracted time dimension features include an average rainfall intensity of 8 mm / h, a maximum rainfall intensity of 15 mm / h, a rate of change of 2, a coefficient of variation of 0.4, and an autocorrelation coefficient of 0.8; the spatial dimension features include a spatial gradient of 2, a clustering degree of 0.75, and a predominant direction from northeast to southwest. After fusion, a fused feature vector containing eight components is formed, comprehensively describing the spatiotemporal characteristics of the rainfall event.

[0041] This invention utilizes multi-sensor data fusion technology to acquire rainfall intensity data in real time and extract fused features, achieving significant technical advantages. By complementing and fusing data from various sensor types, the comprehensiveness and accuracy of rainfall monitoring are improved. Timestamp calibration resolves the issue of inconsistent time signatures across different sensor data, ensuring the continuity and comparability of rainfall evolution sequences. Temporal feature extraction captures the dynamic characteristics of rainfall intensity changes over time, while spatial feature extraction reveals patterns and trends in rainfall distribution across regions. The fusion of these two feature dimensions forms a comprehensive description of rainfall events, providing a rich and reliable information foundation for subsequent rainfall intensity identification and prediction.

[0042] The fused features are matched with the pre-peak evolution sequence features. Based on the historical rules corresponding to the model, the target set of historical rainfall events is selected and the corresponding complete rainfall process features are extracted, including: The fused features and the pre-peak evolution sequence features are used to perform feature mapping calculations to generate a feature similarity matrix, and a rainfall feature matching vector is constructed based on the feature similarity matrix. Based on the mapping rules in the historical rule correspondence model, the similarity of the rainfall feature matching vector is evaluated, the matching degree between each historical rainfall event and the current rainfall event is calculated, and a historical rainfall event matching score is generated. A filtering threshold is set based on the matching score of historical rainfall events, and historical rainfall events with matching scores higher than the filtering threshold are selected to form a target set of historical rainfall events; Based on the order of historical rainfall events in the target historical rainfall event set, a retrieval sequence is constructed by combining the matching score, and the corresponding complete rainfall process features are extracted from the historical rainfall process feature data according to the retrieval sequence.

[0043] Feature mapping is performed between the fused features and the pre-peak evolution sequence features. The fused features are comprehensive features obtained through multi-sensor data fusion, encompassing both temporal rainfall intensity variation and spatial sensor distribution characteristics. The pre-peak evolution sequence features refer to the evolutionary characteristics of rainfall intensity over a period before its peak in historical rainfall events. The feature mapping calculation employs a nonlinear mapping method to transform feature vectors from two different representations into the same feature space for comparison. The nonlinear mapping uses a radial basis function kernel, which can handle nonlinear relationships in high-dimensional feature spaces. During the mapping process, the fused feature vector of the current rainfall event is compared one by one with the pre-peak evolution sequence feature vectors of all rainfall events in the historical database to obtain similarity values. When generating the feature similarity matrix, the rows represent the dimensions of the fused features of the current rainfall event, the columns represent the dimensions of the pre-peak evolution sequence features of historical rainfall events, and the matrix element values ​​are the similarity of the corresponding dimension features. Taking a rainfall event as an example, the current fused feature vector contains 8 dimensions, the historical database contains 500 rainfall events, the pre-peak evolution sequence feature vector of each event contains 10 dimensions, and the generated feature similarity matrix is ​​8×10 in size.

[0044] Constructing a rainfall feature matching vector based on the feature similarity matrix requires comprehensive consideration of the importance and relevance of features across all dimensions. The construction process employs a weighted projection method to compress the two-dimensional similarity matrix into a one-dimensional matching vector. Weight allocation is based on feature importance analysis results, assigning higher weights to feature dimensions that contribute significantly to rainfall identification and lower weights to redundant or noisy features. Weight determination utilizes a learning method based on historical data, obtained by analyzing the correlation between each feature dimension and the rainfall outcome in historical rainfall events. Each element of the rainfall feature matching vector represents the comprehensive similarity between the fused features of the current rainfall event and the pre-peak evolution sequence features of the corresponding rainfall event in the historical database. Taking the aforementioned rainfall event as an example, after weighted projection calculation, a rainfall feature matching vector of length 500 is obtained, where each element's value ranges from 0 to 1, with higher values ​​indicating higher similarity.

[0045] The historical rule-based model contains the feature patterns and evolution laws of different types of rainfall events, stored in the form of a rule set. Mapping rules define the pattern matching conditions and weight allocation methods in the feature space. During similarity evaluation, rainfall feature matching vectors are matched with rules in the historical rule-based model to calculate the degree of matching. The matching degree calculation uses a fuzzy matching algorithm, which can handle the uncertainty and variability of feature values. Different matching strategies are used for different types of rainfall events; for example, for short-duration heavy rainfall, the rate of change of rainfall intensity is emphasized; for continuous rainfall, the cumulative rainfall and duration are emphasized. When calculating the matching degree between each historical rainfall event and the current rainfall event, feature similarity and rule matching degree are considered comprehensively to generate a historical rainfall event matching score. The matching score is calculated using a weighted summation method, with weights determined based on feature importance and rule reliability. Taking a specific rainfall event as an example, matching scores are calculated for 500 historical rainfall events, with scores ranging from 0 to 100; a higher score indicates a higher degree of matching.

[0046] A screening threshold is set based on the matching scores of historical rainfall events. The threshold is determined by the distribution characteristics of the matching scores and the expected number of results. The threshold is dynamically adjusted according to the distribution of the matching scores. When the overall matching scores are high, the threshold is increased to ensure that the most similar events are selected; when the overall matching scores are low, the threshold is appropriately decreased to ensure that there are enough reference events. The threshold setting also considers the characteristics of rainfall types. Higher thresholds are set for rainfall types that are highly typical and have obvious characteristics; relatively lower thresholds are set for complex and variable rainfall types. Historical rainfall events with matching scores higher than the screening threshold are selected from the historical rainfall events to form a target historical rainfall event set. Typically, 10 to 30 historical rainfall events are selected as a reference set. Taking the aforementioned rainfall event as an example, with a screening threshold of 75 points, 20 events with matching scores higher than 75 points are selected from 500 historical rainfall events to form the target historical rainfall event set.

[0047] A retrieval sequence is constructed based on the order of historical rainfall events in the target historical rainfall event set, combined with matching scores. The retrieval sequence is an ordered list containing identifiers of historical rainfall events and their corresponding matching scores, sorted from highest to lowest. The purpose of constructing the retrieval sequence is to effectively organize historical rainfall events, facilitating subsequent retrieval of relevant feature data. Based on the retrieval sequence, corresponding complete rainfall process features are extracted from the historical rainfall process feature data. Complete rainfall process features include the entire lifecycle of rainfall, encompassing all stages from start to finish, especially peak and post-peak features. The extraction process employs a segmented retrieval strategy: first, the corresponding historical event record is located based on the event identifier, and then the complete feature set of that event is extracted. The extracted features are organized chronologically to form a time-series feature set. For each historical rainfall event, the extracted complete rainfall process features include rainfall intensity time series, cumulative rainfall changes, spatial distribution evolution, and other information. Taking a specific target historical rainfall event as an example, the complete rainfall process features include rainfall intensity values ​​at a sampling point every 10 minutes from the start to the end of the rainfall, totaling 24 time points, covering a 4-hour rainfall process, completely recording the entire evolution of rainfall from low to high and then back to low.

[0048] This invention achieves accurate matching between current rainfall events and historical events through feature mapping and similarity assessment, overcoming the matching bias problem caused by relying on single feature comparison. The screening mechanism based on a historical rule-based model improves the intelligence and adaptability of the matching process, enabling differentiated matching strategies for different types of rainfall events. An adaptive threshold setting method ensures the reliability and representativeness of the screening results, avoiding over- or under-screening problems caused by fixed thresholds. By extracting complete rainfall process features, it provides comprehensive and reliable historical reference data for subsequent rainfall intensity prediction.

[0049] A rainfall recurrence prediction model is constructed based on the characteristics of a complete rainfall process. The rainfall intensity evolution process of the current rainfall event is reconstructed through the rainfall recurrence prediction model, including: The complete rainfall process features are decomposed according to the temporal structure to generate a temporal feature mapping vector; Construct a prediction model structure that includes a temporal coding unit, a feature association unit, and a prediction decoding unit; The temporal feature mapping vector is input into the temporal coding unit to extract temporal correlation features; Temporal correlation features are input into the feature association unit to establish a feature mapping relationship, and a prediction parameter matrix is ​​generated based on the feature mapping relationship. The prediction parameter matrix is ​​input into the prediction decoding unit to construct a rainfall recurrence prediction model; The feature sequence of the current rainfall event is input into the rainfall recurrence prediction model. The time-series coding unit generates coded features. The coded features are input into the feature association unit to calculate the feature mapping value. The feature mapping value is input into the prediction decoding unit to generate the rainfall intensity change sequence. The rainfall intensity change sequence is discretized according to the time step to generate a discrete time sequence. The rainfall intensity value corresponding to the discrete time sequence is calculated. The rainfall intensity value is recombined according to the temporal relationship to reconstruct the rainfall intensity evolution process of the current rainfall event.

[0050] The complete rainfall event is decomposed according to its temporal structure to generate a temporal feature mapping vector. The temporal structure decomposition employs a sliding window technique with a window size of 30 minutes and a sliding step size of 10 minutes. For a 4-hour rainfall event, approximately 22 time segments are obtained. For each segment, feature parameters such as average rainfall intensity, rate of change, and peak location are extracted to form a feature sub-vector. Multiple feature sub-vectors are combined chronologically to form the temporal feature mapping vector. Taking a historical rainfall event as an example, the complete rainfall event is decomposed into 22 time segments, and 5 feature parameters are extracted from each segment, generating a temporal feature mapping vector with a dimension of 110.

[0051] A predictive model structure is constructed, comprising a temporal coding unit, a feature association unit, and a prediction decoding unit. The temporal coding unit employs a long short-term memory network structure, including a forget gate, an input gate, and an output gate, with 128 neurons in the hidden layer. The feature association unit uses a self-attention mechanism, containing eight attention heads, each independently learning the association relationships between different feature patterns. The prediction decoding unit employs a fully connected neural network structure, containing three hidden layers with 256, 128, and 64 neurons respectively, using a modified linear unit activation function, and a linear activation function in the output layer.

[0052] The temporal feature mapping vector is input into the temporal coding unit to extract temporal correlation features. The temporal coding unit processes the temporal data through a long short-term memory network to capture the dynamic patterns and trends of rainfall intensity changes. The temporal correlation features contain the dependencies and changing patterns in the time dimension during the rainfall process. Taking the previous example, the input temporal feature mapping vector has a dimension of 110, and the output temporal correlation feature vector has a dimension of 128.

[0053] Temporal correlation features are input into the feature association unit to establish a feature mapping relationship, and a prediction parameter matrix is ​​generated based on the feature mapping relationship. The feature association unit calculates the association weights between features through a self-attention mechanism, and calculates the similarity between features through three projection matrices: query vector, key vector, and value vector, converting them into weight coefficients for weighting the value vector. The multi-head attention layer divides the features into 8 parts, calculates attention for each part, and then merges them. The feature mapping relationship reflects the mutual influence and transformation patterns between rainfall features at different time points. In the example above, the temporal correlation feature dimension is 128, which is processed to generate a prediction parameter matrix with a dimension of 256.

[0054] The prediction parameter matrix is ​​input into the prediction decoding unit to construct a rainfall recurrence prediction model. The prediction decoding unit converts the feature mapping relationship into prediction model parameters using a fully connected neural network, and optimizes them using a gradient descent algorithm with mean squared error as the loss function. The feature sequence of the current rainfall event is input into the rainfall recurrence prediction model. The temporal encoding unit generates coded features, which are then input into the feature association unit to calculate feature mapping values. These feature mapping values ​​are then input into the prediction decoding unit to generate a rainfall intensity change sequence. The feature sequence of the current rainfall event includes rainfall feature data from the most recent 30 minutes to 1 hour. The coded features capture short-term rainfall change patterns, the feature mapping values ​​reflect the degree of matching between the current rainfall features and historical patterns, and the rainfall intensity change sequence represents the future trend of rainfall intensity. Taking a specific rainfall event as an example, inputting the feature sequence of the most recent 60 minutes generates a rainfall intensity change sequence for the next 180 minutes.

[0055] The rainfall intensity variation sequence is discretized according to a time step to generate a discrete time sequence. The rainfall intensity values ​​corresponding to the discrete time sequences are calculated, and these values ​​are reassembled according to their temporal relationship to reconstruct the evolution of rainfall intensity for the current rainfall event. The discretization process uses a time step of 10 minutes or 15 minutes, extracting the intensity values ​​at discrete moments from the continuous sequence and arranging them in chronological order to form a complete sequence. The reconstructed rainfall intensity evolution process includes key information such as peak rainfall intensity, duration, and end time. Using the aforementioned example, a 180-minute continuous sequence is discretized with a 10-minute step, generating predicted values ​​for 19 time points, which are then reassembled into a complete evolution prediction result.

[0056] This invention constructs a rainfall recurrence prediction model based on the characteristics of a complete rainfall process. This model reconstructs the evolution of rainfall intensity in the current rainfall event, demonstrating significant technical effectiveness. Temporal structure decomposition captures the dynamic change patterns during the rainfall process, overcoming the shortcomings of traditional methods in processing temporal information. The temporal coding unit effectively handles long-term dependencies in rainfall data, the feature association unit highlights the influence of key features through a self-attention mechanism, and the prediction decoding unit achieves accurate rainfall intensity reconstruction. Discretization and temporal reassembly techniques ensure the practicality and interpretability of the prediction results.

[0057] like Figure 2 As shown, based on the reconstructed rainfall intensity evolution process, the inflow load change of the urban drainage system is calculated, and a dynamic drainage scheduling scheme is generated, including: The reconstructed complete rainfall intensity evolution process is converted into a time series. The runoff coefficient is calculated based on the land cover data to generate inflow load changes. The drainage system load curve is constructed based on the inflow load changes. Hydraulic calculations are performed on the load curve of the drainage system to obtain the hydraulic parameters of the pipe network nodes. Based on the hydraulic parameters, a pipe network hydraulic state matrix is ​​constructed to generate a drainage capacity assessment value. Based on the drainage capacity assessment value and the pipeline status matrix, the pump station start-up and shutdown parameters and the gate opening parameters are calculated, and the pump station start-up and shutdown parameters and the gate opening parameters are combined to generate the drainage facility operation parameters. The operating parameters of the drainage facilities are divided into multiple time periods according to fixed time intervals. Pump station operation instructions and gate adjustment instructions are generated for each time period to form a time-segmented scheduling instruction sequence. A dynamic drainage scheduling scheme is constructed based on the time-segmented scheduling instruction sequence.

[0058] The reconstructed complete rainfall intensity evolution process was converted into a time series. The runoff coefficient was calculated based on land cover data to generate inflow load changes, and a drainage system load curve was constructed based on these changes. The reconstructed rainfall intensity evolution process consists of discrete rainfall intensity values ​​arranged in chronological order, with a time resolution of 10 minutes. A linear interpolation method was used during the time series conversion to transform the 10-minute resolution data into a 5-minute resolution, enhancing temporal accuracy. Land cover data includes spatial distribution information for different land cover types, categorized as green space, paved surfaces, and roofs, stored in raster form with a resolution of 5 meters. The runoff coefficient was calculated based on land cover type: 0.15 for green space, 0.85 for paved surfaces, and 0.9 for roofs. The inflow load change is the product of rainfall and the runoff coefficient, reflecting the amount of rainwater discharged into the pipe network. The drainage system load curve is a function of the inflow load over time, describing the load state of the drainage system during rainwater discharge. For example, in a rainfall event in a certain area, the reconstructed maximum rainfall intensity is 75 mm / h, the duration is 3 hours, the average runoff coefficient of the area is 0.65, and the calculated peak value of the drainage system load curve is 48.75 m. 3 / s.

[0059] Hydraulic calculations were performed on the load curves of the drainage system to obtain the hydraulic parameters of the pipe network nodes. Based on these parameters, a pipe network hydraulic state matrix was constructed to generate a drainage capacity assessment value. The hydraulic calculations employed a dynamic wave model, considering the unsteady flow characteristics within the pipe network. The explicit finite difference method was used to solve the Saint-Venant equations during the calculations, with a time step of 30 seconds to satisfy stability conditions. The hydraulic parameters of the pipe network nodes included four key indicators: node water level, flow rate, velocity, and pressure. Node water level reflects the water accumulation status, flow rate represents the throughput capacity, velocity reflects the transport efficiency, and pressure is related to the pipe's load-bearing capacity. The pipe network hydraulic state matrix is ​​a multi-dimensional data structure, with each row representing a time step and each column representing a node. The matrix elements are the node hydraulic parameters. The drainage capacity assessment value was calculated based on the pipe network hydraulic state matrix, comprehensively considering factors such as the proportion of node water levels exceeding warning levels, the proportion of full-flow pipes, and the pump station's workload. The assessment value ranged from 0 to 100, with higher values ​​indicating stronger drainage capacity. In the example above, the hydraulically calculated pipe network has 156 nodes, the constructed hydraulic state matrix has a dimension of 360×156×4, and the evaluated drainage capacity value is 72.

[0060] Based on the drainage capacity assessment value and the pipeline network state matrix, pump station start-up and shutdown parameters and gate opening parameters are calculated. These parameters are then combined to generate drainage facility operation parameters. Pump station start-up and shutdown parameters are calculated based on a comparison of node water levels with set thresholds. The pump station starts when the upstream node water level exceeds the start-up threshold and the downstream node water level is below the safety threshold; it stops when the upstream node water level is below the stop threshold or the downstream node water level exceeds the danger threshold. Gate opening parameters are calculated based on the upstream and downstream water level difference and flow demand, with an opening range of 0% to 100%. The drainage facility operation parameters are a combination of pump station start-up and shutdown parameters and gate opening parameters, forming a multi-dimensional control vector to guide the coordinated operation of the drainage facilities. In the aforementioned example, the drainage system includes 5 pump stations and 8 regulating gates. The calculated drainage facility operation parameters are a 13-dimensional control vector, containing 5 binary pump station start-up and shutdown state values ​​and 8 continuous gate opening values.

[0061] The operating parameters of the drainage facilities are divided into multiple time periods according to fixed time intervals. Pump station operation commands and gate adjustment commands are generated for each time period, forming a time-segmented scheduling command sequence. A dynamic drainage scheduling scheme is constructed based on this sequence. The fixed time interval is set to 15 minutes to match the response time of the drainage facilities. Pump station operation commands include three elements: start time, running duration, and stop time. Gate adjustment commands include three elements: adjustment time, opening degree, and adjustment rate. The time-segmented scheduling command sequence is a set of control commands arranged chronologically, covering the entire rainfall process and its subsequent drainage phase. The dynamic drainage scheduling scheme is a complete operation plan built upon the time-segmented scheduling command sequence, containing all the details of facility operation. In the aforementioned example, the rainfall lasted for 3 hours, with a subsequent drainage phase of 2 hours, divided into 20 time periods. The generated time-segmented scheduling command sequence contains 100 basic commands. The constructed dynamic drainage scheduling scheme can automatically adapt to changes in rainfall intensity, achieving intelligent and coordinated control of the drainage facilities.

[0062] This invention calculates the inflow load changes of urban drainage systems based on the reconstructed rainfall intensity evolution process, generating a dynamic drainage scheduling scheme. By accurately transforming the rainfall intensity evolution process and combining it with land cover data to calculate the runoff coefficient, accurate prediction of inflow load is achieved. Based on hydraulic state and drainage capacity assessment, precise pump station start-up and shutdown parameters and gate opening parameters are calculated, enabling coordinated control of drainage facilities. Through time-segmented scheduling command sequences, a dynamic drainage scheduling scheme with rapid response and strong adaptability is constructed.

[0063] The multi-sensor data fusion-based rapid rainfall intensity identification system provided in this embodiment of the invention includes: The historical data processing module is used to acquire historical precipitation datasets for the study area, plot peak curves of rainfall intensity, segment rainfall events based on peak curves, and obtain pre-peak evolution sequence characteristics. The rule modeling module is used to perform correlation analysis between the pre-peak evolution sequence features and the complete rainfall process features to establish a historical rule correspondence model. The complete rainfall process features include pre-peak evolution features, peak features, and post-peak decay features. The real-time monitoring module includes multiple rainfall sensors, which are used to collect rainfall intensity data of the current rainfall event in real time, generate the current rainfall evolution sequence, and extract the rainfall intensity change features in the time dimension and the sensor distribution features in the spatial dimension to form fused features; The feature matching module is used to match the fused features with the pre-peak evolution sequence features, filter out the target set of historical rainfall events according to the historical rule corresponding model, and extract the corresponding complete rainfall process features. The prediction modeling module is used to construct a rainfall recurrence prediction model based on the characteristics of the complete rainfall process, and to reconstruct the rainfall intensity evolution process of the current rainfall event through the rainfall recurrence prediction model; The scheduling scheme generation module is used to calculate the inflow load changes of the urban drainage system based on the reconstructed rainfall intensity evolution process and generate a dynamic drainage scheduling scheme.

[0064] One technical solution provided in this embodiment of the invention is an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps in any of the aforementioned methods.

[0065] One technical solution provided in this embodiment of the invention is a computer-readable storage medium storing a computer program, wherein the processor executes the computer program to implement the steps in any of the aforementioned methods.

[0066] The specific embodiments described above are preferred embodiments of the present invention and are not intended to limit the specific scope of the present invention. The scope of the present invention includes, but is not limited to, these specific embodiments. All equivalent changes made in accordance with the shape and structure of the present invention are within the protection scope of the present invention.

Claims

1. A method for rapid identification of rainfall intensity based on multi-sensor data fusion, characterized in that, Includes the following steps: Historical precipitation datasets for the study area are obtained, peak precipitation intensity curves are plotted, and precipitation events are segmented based on the peak precipitation curves to obtain pre-peak evolution sequence characteristics. A correlation analysis was conducted between the pre-peak evolution sequence characteristics and the characteristics of the complete rainfall process to establish a historical rule correspondence model. The characteristics of the complete rainfall process include pre-peak evolution characteristics, peak characteristics, and post-peak attenuation characteristics. The system collects rainfall intensity data of the current rainfall event in real time by multiple rainfall sensors, generates the current rainfall evolution sequence, and extracts the rainfall intensity change features in the time dimension and the sensor distribution features in the spatial dimension to form a fused feature. The fused features are matched with the pre-peak evolution sequence features, and the target set of historical rainfall events is selected according to the historical rule corresponding model and the corresponding complete rainfall process features are extracted. A rainfall recurrence prediction model is constructed based on the characteristics of a complete rainfall process, and the rainfall intensity evolution process of the current rainfall event is reconstructed through the rainfall recurrence prediction model. Based on the reconstructed rainfall intensity evolution process, the inflow load change of the urban drainage system is calculated, and a dynamic drainage scheduling scheme is generated.

2. The method according to claim 1, characterized in that, Historical precipitation datasets for the study area were obtained, and peak precipitation intensity curves were plotted. Based on the peak precipitation curves, precipitation events were segmented to obtain pre-peak evolution sequence characteristics, including: Perform time series analysis on historical precipitation datasets, construct an adaptive sliding window, calculate the mean and variance of precipitation intensity data within the adaptive sliding window, generate a peak identification threshold, and use the peak identification threshold to identify local maxima of precipitation intensity within the adaptive sliding window. Calculate the time interval and intensity variation between local maxima of rainfall intensity to construct a state variable matrix; The significance index is calculated based on the state quantity matrix, and the local maxima of rainfall intensity are sorted. The maxima with significance index higher than the mean are selected as the rainfall peak points, and the rainfall peak points are connected to form the rainfall intensity peak curve. The fluctuation trend of the peak rainfall intensity curve is analyzed, and the fluctuation inflection point position is extracted from the peak rainfall intensity curve. The fluctuation inflection point position is used as the segment boundary point of the rainfall event, and the pre-peak evolution sequence features of the rainfall event are extracted based on the segment boundary point.

3. The method according to claim 1, characterized in that, The correlation analysis between the pre-peak evolution sequence characteristics and the characteristics of the complete rainfall process was used to establish a historical rule correspondence model, including: Extract the rainfall feature vector from the pre-peak evolution sequence features, retrieve the corresponding time period feature vector of the rainfall feature vector in the complete rainfall process features, and calculate the similarity value between the rainfall feature vector and the corresponding time period feature vector; Historical rainfall events are grouped based on similarity values, the state transition probability of rainfall events within each group is calculated, and the state transition matrix of rainfall events is constructed using the state transition probability. Extract the state transition patterns from the state transition matrix of rainfall events and construct a set of state transition rules; The rules in the state transition rule set are subjected to reliability assessment, a rule assessment score is generated, transition rules with assessment scores higher than a preset threshold are selected, and a historical rule correspondence model is established.

4. The method according to claim 1, characterized in that, Rainfall intensity data for the current rainfall event is collected in real time by multiple rainfall sensors, generating the current rainfall evolution sequence. The temporal dimensions of rainfall intensity variation and the spatial dimensions of sensor distribution characteristics are extracted to form fused features, including: A timestamp calibration operation is performed on the rainfall intensity data collected by the rainfall sensor to generate calibrated rainfall intensity data. Based on the calibrated rainfall intensity data, a current rainfall evolution sequence is generated, and the time dimension rainfall intensity change features are extracted from the current rainfall evolution sequence. The spatial coordinates of the rainfall sensor are obtained, and the calibrated rainfall intensity data is combined with the spatial coordinates to construct a spatial distribution matrix. The sensor distribution features in the spatial dimension are extracted from the spatial distribution matrix. The time-dimensional rainfall intensity variation characteristics and the spatial sensor distribution characteristics are fused to generate fused features.

5. The method according to claim 1, characterized in that, The fused features are matched with the pre-peak evolution sequence features. Based on the historical rules corresponding to the model, the target set of historical rainfall events is selected and the corresponding complete rainfall process features are extracted, including: The fused features and the pre-peak evolution sequence features are used to perform feature mapping calculations to generate a feature similarity matrix, and a rainfall feature matching vector is constructed based on the feature similarity matrix. Based on the mapping rules in the historical rule correspondence model, the similarity of the rainfall feature matching vector is evaluated, the matching degree between each historical rainfall event and the current rainfall event is calculated, and a historical rainfall event matching score is generated. A filtering threshold is set based on the matching score of historical rainfall events, and historical rainfall events with matching scores higher than the filtering threshold are selected to form a target set of historical rainfall events; Based on the order of historical rainfall events in the target historical rainfall event set, a retrieval sequence is constructed by combining the matching score, and the corresponding complete rainfall process features are extracted from the historical rainfall process feature data according to the retrieval sequence.

6. The method according to claim 1, characterized in that, A rainfall recurrence prediction model is constructed based on the characteristics of a complete rainfall process. The rainfall intensity evolution process of the current rainfall event is reconstructed through the rainfall recurrence prediction model, including: The complete rainfall process features are decomposed according to the temporal structure to generate a temporal feature mapping vector; Construct a prediction model structure that includes a temporal coding unit, a feature association unit, and a prediction decoding unit; The temporal feature mapping vector is input into the temporal coding unit to extract temporal correlation features; Temporal correlation features are input into the feature association unit to establish a feature mapping relationship, and a prediction parameter matrix is ​​generated based on the feature mapping relationship. The prediction parameter matrix is ​​input into the prediction decoding unit to construct a rainfall recurrence prediction model; The feature sequence of the current rainfall event is input into the rainfall recurrence prediction model. The time-series coding unit generates coded features. The coded features are input into the feature association unit to calculate the feature mapping value. The feature mapping value is input into the prediction decoding unit to generate the rainfall intensity change sequence. The rainfall intensity change sequence is discretized according to the time step to generate a discrete time sequence. The rainfall intensity value corresponding to the discrete time sequence is calculated. The rainfall intensity value is recombined according to the temporal relationship to reconstruct the rainfall intensity evolution process of the current rainfall event.

7. The method according to claim 1, characterized in that, Based on the reconstructed rainfall intensity evolution, the inflow load changes of the urban drainage system are calculated, and a dynamic drainage scheduling scheme is generated, including: The reconstructed complete rainfall intensity evolution process is converted into a time series. The runoff coefficient is calculated based on the land cover data to generate inflow load changes. The drainage system load curve is constructed based on the inflow load changes. Hydraulic calculations are performed on the load curve of the drainage system to obtain the hydraulic parameters of the pipe network nodes. Based on the hydraulic parameters, a pipe network hydraulic state matrix is ​​constructed to generate a drainage capacity assessment value. Based on the drainage capacity assessment value and the pipeline status matrix, the pump station start-up and shutdown parameters and the gate opening parameters are calculated, and the pump station start-up and shutdown parameters and the gate opening parameters are combined to generate the drainage facility operation parameters. The operating parameters of the drainage facilities are divided into multiple time periods according to fixed time intervals. Pump station operation instructions and gate adjustment instructions are generated for each time period to form a time-segmented scheduling instruction sequence. A dynamic drainage scheduling scheme is constructed based on the time-segmented scheduling instruction sequence.

8. A rapid rainfall intensity identification system based on multi-sensor data fusion, used to implement the method described in any one of claims 1-7, characterized in that, The system includes: The historical data processing module is used to acquire historical precipitation datasets for the study area, plot peak curves of rainfall intensity, segment rainfall events based on peak curves, and obtain pre-peak evolution sequence characteristics. The rule modeling module is used to perform correlation analysis between the pre-peak evolution sequence features and the complete rainfall process features to establish a historical rule correspondence model. The complete rainfall process features include pre-peak evolution features, peak features, and post-peak decay features. The real-time monitoring module includes multiple rainfall sensors, which are used to collect rainfall intensity data of the current rainfall event in real time, generate the current rainfall evolution sequence, and extract the rainfall intensity change features in the time dimension and the sensor distribution features in the spatial dimension to form fused features; The feature matching module is used to match the fused features with the pre-peak evolution sequence features, filter out the target set of historical rainfall events according to the historical rule corresponding model, and extract the corresponding complete rainfall process features. The prediction modeling module is used to construct a rainfall recurrence prediction model based on the characteristics of the complete rainfall process, and to reconstruct the rainfall intensity evolution process of the current rainfall event through the rainfall recurrence prediction model; The scheduling scheme generation module is used to calculate the inflow load changes of the urban drainage system based on the reconstructed rainfall intensity evolution process and generate a dynamic drainage scheduling scheme.

9. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 7.

Images