CN-122020009-A - Intelligent channel water depth prediction method and system based on multi-source information collaborative fusion

CN122020009ACN 122020009 ACN122020009 ACN 122020009ACN-122020009-A

Abstract

The invention discloses an intelligent channel water depth prediction method and system based on multi-source information collaborative fusion, wherein the method comprises the steps of processing acquired original multi-source observation data to obtain input data with a unified space-time structure; the fused data is input into a trained deep learning water depth prediction model to obtain a minute-level dynamic prediction result of the water depth of the channel, wherein the water depth prediction model is a time sequence feature extraction network comprising a self-attention mechanism and a long-short-period memory unit, a transducer structure is fused to extract long-distance dependency relations and is combined with an LSTM unit to capture short-term dynamic change trends, and a monotonic water head operator layer is introduced into the model structure to ensure that predicted output meets a preset physical monotonic constraint relation. The method can realize minute-level dynamic water depth prediction based on the constructed deep learning water depth prediction model through collaborative fusion of multi-source observation data, and has the advantages of high prediction precision, stability, interpretability and the like.

Inventors

XU SUDONG
XU HAIYANG
ZHANG NINI
Li Boju
LU XIAODONG
YIN JIE
MA MENGDI
WANG ZONGCHUAN
MAO LIUYAN
TAN WEIKAI
TANG SHUANG

Assignees

东南大学

Dates

Publication Date: 20260512
Application Date: 20251203

Claims (10)

1. An intelligent channel water depth prediction method based on multi-source information collaborative fusion is characterized by comprising the following steps: Processing the collected original multi-source observation data to obtain input data with a unified space-time structure; the fused data is input into a trained deep learning water depth prediction model to obtain a minute-level dynamic prediction result of the water depth of the channel, wherein the water depth prediction model is a time sequence feature extraction network comprising a self-attention mechanism and a long-short-period memory unit, a transducer structure is fused to extract long-distance dependency relations and is combined with an LSTM unit to capture short-term dynamic change trends, and a monotonic water head operator layer is introduced into the model structure to ensure that predicted output meets a preset physical monotonic constraint relation.
2. The intelligent channel water depth prediction method based on multi-source information collaborative fusion according to claim 1, wherein the obtaining mode of the original multi-source observation data comprises the following steps: Arranging multi-source observation equipment along the typical river reach of a channel, wherein the observation equipment comprises a shore fixed radar fluviograph, a water surface floating type or shipborne multi-beam depth sounder, a hydrological monitoring system carried by an unmanned survey ship, a remote sensing satellite or unmanned aerial vehicle imaging device, a ground automatic weather station and a GNSS high-precision positioning module; the multi-source observation data is synchronously uploaded to a system data center through wireless communication or a local storage mode, and timestamp marking, error checking and preliminary cleaning are carried out in a data processing unit to form a structured original input data set.
3. The intelligent channel water depth prediction method based on multi-source information collaborative fusion according to claim 1 is characterized in that the processing of multi-source observation data comprises time synchronization processing, kalman filtering denoising, outlier detection and missing value interpolation, the input of fused data into a deep learning water depth prediction model comprises the steps of completing weighted fusion according to historical confidence of each channel, combining prior and likelihood to carry out Bayesian calibration, constructing derived features by identifying mutation and gate action events, including difference, change rate and smoothness, calculating autocorrelation and cross correlation to describe time sequence dependence and cross source coupling, unifying dimension and standardization, carrying out time coding, slicing according to a fixed time window and applying causal mask to form a space-time sequence which can be used for modeling, reserving upstream and downstream water levels and gate opening and closing states to a previous layer in a straight-through mode, and carrying out directivity correction on a monotonic water head operator layer.
4. The intelligent channel water depth prediction method based on the multi-source information collaborative fusion is characterized in that the input of a deep learning water depth prediction model is fused multi-source space-time observation data, the input of the deep learning water depth prediction model is a minute-level prediction value of channel water depth, training samples comprise historical multi-source space-time observation data and channel water depth actual measurement values corresponding to the historical multi-source space-time observation data, the deep learning water depth prediction model comprises a monotonic water head operator layer, the monotonic water head operator layer is arranged at the output end of a transducer-LSTM network and is positioned in front of a final output layer and used for applying monotonic constraint on the influence of key physical variables, the key physical variables comprise upstream water level and gate opening degree as monotonic non-decreasing variables, and downstream water level and gate closing degree as monotonic non-increasing variables.
5. The intelligent channel water depth prediction method based on multi-source information collaborative fusion according to claim 4, wherein the training samples are obtained by the following steps: Based on a continuous multi-source observation data sequence and a corresponding channel water depth actual measurement value, carrying out time synchronization and alignment on data at each observation time, and detecting and filling missing data to obtain multi-source data characteristics at each observation time and a corresponding channel water depth actual measurement value; the method comprises the steps of constructing continuous observation data in each time window into a model input sequence by dividing according to the length of a preset time window; Based on the time period covered by each input sequence, extracting the channel water depth actual measurement value corresponding to the tail moment of the sequence, and taking the channel water depth actual measurement value as a prediction output label of the input sequence, wherein the prediction output label is used for supervising the training target of a model in learning, and ensuring that each input data corresponds to a unique output result so as to construct a standard input-output sample pair; and (3) through traversing all observation sequences, acquiring complete input data and output tag pairs, and constructing and forming a sample set for training.
6. The intelligent channel water depth prediction method based on multi-source information collaborative fusion according to claim 1 is characterized in that a mixed neural network architecture is adopted in the deep learning water depth prediction model, dynamic prediction of channel water depth is achieved through stacking a transducer network and an LSTM network, the transducer network part comprises two encoding layers and two feedforward neural network layers based on a multi-head self-attention mechanism from input to output, long-distance dependency features of a time sequence are extracted through a residual connection and regularization unit, the LSTM network part comprises an input layer, two hidden layers and an output layer from input to output, the input layer receives the time sequence features extracted by the transducer network part, primary prediction representation of a main network is obtained through extraction of short-term dynamic change features by the two hidden layers, a monotonic water head operator layer is simultaneously received before the final output layer, directivity residual errors are generated, increase and decrease of physical consistency is carried out on the primary prediction, and accordingly, the final prediction result meeting monotonic constraint is output, and the channel water depth prediction result is obtained in the final output layer in series.
7. The intelligent channel water depth prediction and multisource information collaborative fusion method is characterized in that a monotonic water head operator layer is arranged in front of a final output layer, implicit representation of a main network and key physical variables are mapped in a combined mode to form directional residual correction, ascending branches of the monotonic water head operator layer are used for reflecting forward influences on water depth when the upstream water level and the opening degree of a gate are increased, descending branches are used for reflecting inhibition influences on water depth when the downstream water level and the closing degree of the gate are increased, residual fusion is finally carried out on the monotonic water head operator layer and a main prediction result, and a minute-level water depth predicted value meeting a physical monotonic relation is output.
8. The intelligent channel water depth prediction and multisource information collaborative fusion method is characterized in that training of a depth learning water depth prediction model is optimized by adopting a joint loss function comprising a plurality of constraints, the joint loss comprises an accuracy item approaching actual measurement water depth, a monotonic consistency item restraining monotonic water head operator layer output to meet a physical monotonic relation and a residual amplitude constraint item used for restraining excessive correction, the monotonic consistency item penalizes the situation that the upstream water level and the gate opening are larger but the prediction result is shallower by constructing a sample pair with only the difference in the magnitude of the upstream water level or the gate opening and the remaining conditions kept consistent in a training set, meanwhile penalizes the situation that the downstream water level or the gate closing degree is larger but the prediction result is deeper by only the sample pair with the difference in the downstream water level and the gate closing degree, and the residual amplitude constraint item limits the correction amplitude of the monotonic water head operator layer to the main prediction result and avoids excessive transition caused by single point abnormality under extreme working conditions.
9. An intelligent channel water depth prediction system based on multi-source information collaborative fusion, comprising a memory, a processor and a computer program stored on the memory and capable of running on the processor, wherein the computer program when executed by the processor realizes the steps of the intelligent channel water depth prediction method based on multi-source information collaborative fusion according to any one of claims 1-8.
10. A computer program product comprising a computer program, characterized in that the computer program, when being executed by a processor, implements the steps of the intelligent channel water depth prediction method based on collaborative fusion of multi-source information according to any one of claims 1-8.

Description

Intelligent channel water depth prediction method and system based on multi-source information collaborative fusion Technical Field The invention relates to channel engineering and multi-source data processing, in particular to a method for realizing real-time prediction of the depth of water in a inland channel based on a deep learning model. Background Along with the continuous improvement of the modernization level of inland water transportation, the demands for finely managing and controlling the navigation channels are increasing, the water depth of the navigation channels is taken as a key parameter for determining the navigation safety of ships, the prediction accuracy of the navigation channels is directly related to the navigation efficiency and the operation safety, and the higher requirements are also put forward for the operation and maintenance of the navigation channels. The traditional channel water depth acquisition mode mostly depends on traditional means such as manual inspection and underwater sounding instruments, and the like, and has the problems of poor real-time performance, limited coverage, large weather and terrain restriction and the like, and is difficult to meet the dynamic water depth response requirements of complex river reach, abrupt weather or high-frequency navigation scenes. Particularly in the inland river section with frequent tidal estuary and sediment accumulation, the water depth is severely changed, and the water depth information is easily delayed due to the dependence on static measurement data, so that the ship scheduling and navigation safety are affected. Disclosure of Invention Aiming at the defect of acquisition of the water depth of the existing channel, the invention aims to provide the intelligent channel water depth prediction method and system based on the collaborative fusion of the multi-source information, which can realize minute-level dynamic water depth prediction and have the advantages of high prediction precision, stability, interpretability and the like. The technical scheme is that in order to achieve the aim of the invention, the invention adopts the following technical scheme: an intelligent channel water depth prediction method based on multi-source information collaborative fusion comprises the following steps: Processing the collected original multi-source observation data to obtain input data with a unified space-time structure; the fused data is input into a trained deep learning water depth prediction model to obtain a minute-level dynamic prediction result of the water depth of the channel, wherein the water depth prediction model is a time sequence feature extraction network comprising a self-attention mechanism and a long-short-period memory unit, a transducer structure is fused to extract long-distance dependency relations and is combined with an LSTM unit to capture short-term dynamic change trends, and a monotonic water head operator layer is introduced into the model structure to ensure that predicted output meets a preset physical monotonic constraint relation. Preferably, the acquiring manner of the original multi-source observation data includes: Arranging multi-source observation equipment along the typical river reach of a channel, wherein the observation equipment comprises a shore fixed radar fluviograph, a water surface floating type or shipborne multi-beam depth sounder, a hydrological monitoring system carried by an unmanned survey ship, a remote sensing satellite or unmanned aerial vehicle imaging device, a ground automatic weather station and a GNSS high-precision positioning module; the multi-source observation data is synchronously uploaded to a system data center through wireless communication or a local storage mode, and timestamp marking, error checking and preliminary cleaning are carried out in a data processing unit to form a structured original input data set. The method comprises the steps of processing multisource observation data, namely time synchronization processing, kalman filtering denoising, outlier detection and missing value interpolation, inputting the fused data into a deep learning water depth prediction model, completing weighted fusion according to historical confidence of each channel, combining priori with likelihood to carry out Bayesian calibration, constructing derivative features by identifying mutation and gate action events, comprising difference, change rate and smoothing, calculating autocorrelation and cross correlation to describe time sequence dependence and cross source coupling, unifying dimension and standardization, carrying out time coding, slicing according to fixed time windows, applying causal masks to form a time-space sequence which can be used for modeling, reserving upstream and downstream water levels and gate opening and closing states to an output previous layer in a straight-through mode, and carrying out directivity correction on a monotonic water head operator