Irrigation area sediment deposition prediction method and system, computer equipment and storage medium

Abstract

The invention provides an irrigation area sediment deposition prediction method and system, computer equipment and a storage medium, and belongs to the field of water conservancy project intellectualization, and the method comprises the steps: collecting historical hydrological data, real-time monitoring data and remote sensing observation data of a to-be-predicted irrigation area; the method comprises the following steps: resampling historical hydrological data and real-time monitoring data, carrying out forward filling completion on missing data, filling remote sensing observation data into a unified time axis, and filling missing spatial features based on spatial feature values of an adjacent time reference to obtain multi-source data with time-space alignment; extracting features of the multi-source data to obtain four features; and combining the four features and calculating correlation weights among the features so as to determine the sediment flow, the sediment concentration and the sediment flow velocity of the to-be-predicted irrigation area. According to the method, aiming at the defect of non-uniform spatial and temporal distribution among multi-source data sources, spatial and temporal alignment is realized through time re-adoption and spatial feature completion, the input information can accurately reflect the sediment transportation process of an irrigation area, and the prediction accuracy is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of intelligentization of water conservancy projects, and particularly relates to a farmland sediment deposition prediction method and system, a computer device and a storage medium. BACKGROUND

[0002] Farmland sediment deposition refers to the phenomenon that the sediment carried by the water flow in the irrigation channel is deposited due to the slowing down of the flow rate. Long-term deposition can significantly reduce the channel water capacity and weaken the irrigation efficiency, and even cause local flood risk. In order to avoid the impact of farmland sediment deposition on people's production and life, the traditional way is to predict based on the experience of water conservancy experts, and to provide flow adjustment, stage sediment discharge and other strategies. However, the expert experience is difficult to quantify the complex water-sediment interaction, and is easily affected by subjective judgment bias, resulting in frequent engineering decision-making errors such as gate scheduling lag and misjudgment of sediment discharge time. Therefore, the perception and prediction problem of farmland sediment deposition needs to be solved systematically.

[0003] In view of the above problems, the existing technology often uses statistical models such as neural networks for intelligent calculation and prediction, but the existing farmland sediment deposition prediction method generally relies on a single data source, which is difficult to accurately reflect the complex sediment transport process in the farmland, affecting the prediction accuracy. SUMMARY

[0004] In order to solve the problem that the existing farmland sediment deposition prediction method relies on a single data source and has low prediction accuracy, the application provides a farmland sediment deposition prediction method, system, computer device and storage medium.

[0005] In order to achieve the above purpose, the application provides the following technical solutions: The farmland sediment deposition prediction method comprises: Collecting historical hydrological data, real-time monitoring data and remote sensing observation data of a to-be-predicted farmland; resampling the historical hydrological data and the real-time monitoring data according to a unified time axis of the to-be-predicted farmland, filling the missing historical hydrological data or real-time monitoring data in the unified time axis through forward filling, filling the remote sensing observation data into the unified time axis, and filling the missing spatial features in the remote sensing observation data based on the spatial feature values of the adjacent time reference on the unified time axis to obtain spatio-temporally aligned multi-source data, wherein the unified time axis covers the whole period of the to-be-predicted farmland; Extracting features from the spatio-temporally aligned multi-source data to obtain farmland engineering morphology features, farmland hydrological index statistical features, farmland spatial grid features and farmland time series operation features; The irrigation area engineering morphological characteristics, irrigation area hydrological index statistical characteristics, irrigation area spatial grid characteristics and irrigation area time sequence operation characteristics are combined, and the correlation weights between the characteristics are calculated based on a multi-head attention mechanism to obtain an attention feature matrix, and the sediment flow, sediment concentration and sediment flow rate of the to-be-predicted irrigation area are determined according to the attention feature matrix.

[0006] The application further provides an irrigation area sediment deposition prediction system, comprising: An irrigation area data acquisition module is configured to acquire historical hydrological data, real-time monitoring data and remote sensing observation data of a to-be-predicted irrigation area, resample the historical hydrological data and real-time monitoring data according to a unified time axis of the to-be-predicted irrigation area, complete the missing historical hydrological data or real-time monitoring data in the unified time axis by forward filling, fill the remote sensing observation data into the unified time axis, and fill the missing spatial characteristics in the remote sensing observation data based on the spatial characteristic values of adjacent time references on the unified time axis to obtain multi-source data aligned in time and space, wherein the unified time axis covers the whole period of the to-be-predicted irrigation area. A multi-source feature extraction module is configured to extract features from the multi-source data aligned in time and space to obtain irrigation area engineering morphological characteristics, irrigation area hydrological index statistical characteristics, irrigation area spatial grid characteristics and irrigation area time sequence operation characteristics. A sediment deposition prediction module is configured to combine the irrigation area engineering morphological characteristics, irrigation area hydrological index statistical characteristics, irrigation area spatial grid characteristics and irrigation area time sequence operation characteristics, and calculate the correlation weights between the characteristics based on a multi-head attention mechanism to obtain an attention feature matrix, and determine the sediment flow, sediment concentration and sediment flow rate of the to-be-predicted irrigation area according to the attention feature matrix.

[0007] The application further provides a computer device comprising a memory, a processor and a computer program stored on the memory, wherein the processor executes the computer program to implement the steps of any one of the irrigation area sediment deposition prediction methods.

[0008] The application further provides a computer readable storage medium, wherein the storage medium stores a computer program, and the computer program can implement the steps of any one of the irrigation area sediment deposition prediction methods when loaded by a processor.

[0009] The irrigation area sediment deposition prediction method provided by the application has the following beneficial effects: The irrigation area silt accumulation prediction method provided by the application considers the limitation of a single data source, collects historical hydrological data, real-time monitoring data and remote sensing observation data of the irrigation area to be predicted as multi-source data for feature extraction in the model, and fills in the data lacking in the historical hydrological data, real-time monitoring data and remote sensing observation data in a forward direction and fills in the characteristic values of adjacent time bases, so that the irrigation area engineering form features, irrigation area hydrological index statistical features, irrigation area spatial grid features and irrigation area time sequence operation features can fully reflect the overall silt transport process of the irrigation area, and the four types of features are fused for silt accumulation prediction, thereby improving the accuracy of the irrigation area silt accumulation prediction.

[0010] In addition, considering that there is a problem of too large deviation in direct prediction due to the uneven spatio-temporal distribution among the multi-source data, the irrigation area silt accumulation prediction method provided by the application resamples the historical hydrological data and real-time monitoring data according to a unified time axis, and fills in the remote sensing observation data into the unified time axis, so that the spatio-temporal alignment among the multi-source data is realized, and the influence of uneven spatio-temporal distribution of data on silt accumulation prediction is avoided. BRIEF DESCRIPTION OF DRAWINGS

[0011] In order to more clearly illustrate the embodiments of the application and the design scheme thereof, the drawings required by the embodiments will be briefly introduced as follows. The drawings in the following description are only part of the embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0012] Figure 1 Fig. 1 is a schematic diagram of the irrigation area silt accumulation prediction method provided by the embodiment of the application; Figure 2 Fig. 2 is a schematic diagram of the multi-feature three-dimensional spatial distribution provided by the embodiment of the application; Figure 3 Fig. 3 is a schematic diagram of the irrigation area silt accumulation prediction method provided by the embodiment of the application; Figure 4 Fig. 4 is a schematic diagram of the irrigation area silt accumulation prediction method provided by the embodiment of the application; Figure 5 Fig. 5 is a schematic diagram of the deep learning model structure provided by the embodiment of the application; Figure 6 Fig. 6 is a schematic diagram of the irrigation area silt accumulation prediction method provided by the embodiment of the application; Figure 7 Fig. 7 is a schematic diagram of the irrigation area silt accumulation prediction method provided by the embodiment of the application; Figure 8 Fig. 8 is a flow chart of intelligent decision-making and explainability analysis provided by the embodiment of the application; Figure 9A flowchart of a silt accumulation prediction technology process for an embodiment of the present application is provided. DETAILED DESCRIPTION

[0013] In order for those skilled in the art to better understand the technical solutions of the present application and to implement them, the present application will be described in detail below in conjunction with the drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present application, and cannot be used to limit the protection scope of the present application.

[0014] In view of the defects of the prior art, the present application provides a silt accumulation prediction method for an irrigation area, a multi-source heterogeneous data space-time alignment method, which can efficiently fuse multiple types of data such as historical monitoring, real-time monitoring and remote sensing, realize unified space-time indexing and intelligent filling of missing values of the data, and through automatic extraction and fusion of multi-dimensional features such as time, space and physics, combined with hydraulic and sediment dynamics parameters such as Manning coefficient, hydraulic radius and Froude number, a high-expression feature system of physical-data fusion is constructed, which can fully reflect the complex silt transport process of the irrigation area, and provide a comprehensive data basis for model prediction.

[0015] Furthermore, the silt accumulation prediction method provided by the present application sets a SHAP explainability analysis tool in the model, realizes feature importance analysis and causal relationship chain modeling of the model prediction results. The system can automatically reveal the influence path of various features on the silt accumulation risk, support causal inference and decision transparency, greatly improve the credibility and engineering usability of the model, and provide scientific and traceable decision basis for water conservancy managers.

[0016] In addition, the silt accumulation prediction method provided by the present application constructs an intelligent display platform integrating multiple dynamic visualization functions such as risk thermodynamic map, feature importance map and prediction animation, which supports intuitive presentation of the spatio-temporal evolution process of the silt accumulation risk. The system has good generalization ability and adaptability, can automatically adapt to different irrigation areas and variable hydro-meteorological conditions, supports real-time data access and intelligent scheduling strategy generation, and provides efficient and scalable technical support for intelligent management and risk warning of digital twin irrigation areas.

[0017] Embodiment 1 The present application provides a silt accumulation prediction method for an irrigation area, as shown in Figure 1 The method comprises the following steps: Step 11, collect historical hydrological data, real-time monitoring data and remote sensing observation data of the irrigation area to be predicted; resample the historical hydrological data and real-time monitoring data according to the unified time axis of the irrigation area to be predicted, fill in the missing historical hydrological data or real-time monitoring data through forward filling, fill in the remote sensing observation data into the unified time axis, fill in the missing spatial features of the remote sensing observation data based on the spatial feature values of the adjacent time reference on the unified time axis, and obtain the spatio-temporal aligned multi-source data, wherein the unified time axis covers the whole period of the irrigation area to be predicted.

[0018] Step 12, feature extraction is performed on the spatio-temporal aligned multi-source data to obtain irrigation area engineering morphological features, irrigation area hydrological index statistical features, irrigation area spatial grid features and irrigation area time sequence operation features.

[0019] Step 13, combine the irrigation area engineering morphological features, irrigation area hydrological index statistical features, irrigation area spatial grid features and irrigation area time sequence operation features, and calculate the correlation weight between the features based on the multi-head attention mechanism to obtain an attention feature matrix, and determine the sediment flow, sediment concentration and sediment flow rate of the irrigation area to be predicted according to the attention feature matrix.

[0020] Specifically, in the irrigation area sediment deposition prediction method, first, collect multi-source heterogeneous data such as historical hydrological data, real-time monitoring data and remote sensing observation data of the irrigation area to be predicted, and perform data format identification and data value confirmation, for example, preliminary adjustment of missing values and abnormal values, and through a space-time alignment processor, the multi-source heterogeneous data is uniformly time resampled and spatial feature completed, for example, through forward filling and backward filling, and spatial interpolation method to complete the missing data and enhance the multi-source data. In addition, the multi-source heterogeneous data can also be calculated for complex spatial features such as cross-section spatial relationship and upstream influence, so as to construct a spatio-temporally aligned data set.

[0021] Then, feature extraction and feature enhancement are performed on the spatio-temporally aligned data set to obtain, for example, irrigation area physical features such as irrigation area flow, irrigation area water depth, sediment concentration, flow rate, river width, curvature, irrigation area statistical features such as sliding mean, standard deviation, extreme value, change rate, irrigation area spatial features such as cross-section distance, upstream flow influence, spatial lag feature, spatial correlation, and irrigation area time features such as hour, day, month, season, holiday, and the like, which represent the irrigation area time sequence operation, and the irrigation area physical features, irrigation area statistical features, irrigation area spatial features and irrigation area time features are standardized to ensure that the data quality meets the model input requirements.

[0022] Finally, the sediment prediction is performed by a pre-trained multi-layer deep neural network model to generate a sediment deposition prediction result. The multi-layer deep neural network model is constructed based on a Transformer structure, and 52-dimensional fusion features corresponding to physical features, statistical features, spatial features and temporal features of the irrigation area are input to the model. The model generates multi-target prediction results of sediment concentration, sediment flow rate and sediment flow rate of the to-be-identified irrigation area through an encoder, a decoder and a position encoder, and the Nash efficiency coefficients (NSE) of the three indexes of sediment concentration, sediment flow rate and sediment flow rate are all greater than 0.97. The optimal weight file corresponding to the optimal weight of the multi-target is determined based on the loss function and the accuracy index during the model training process. For example, during the model training process, the model automatically records the loss function, various accuracy indexes, and generates a training process visualization report. When the model training is completed, the model automatically saves the optimal weight file to facilitate subsequent inference and deployment.

[0023] The 52-dimensional fusion features are physical and data fusion features, which include flow, water depth, sediment concentration, flow rate, river width, curvature, and other core physical parameters. For example, the deep coupling of physical laws and data features is realized by the hydraulic radius calculation formula shown in formula (1), the Froude number calculation formula shown in formula (2), and the Reynolds number calculation formula shown in formula (3):

[0024] R=A / P (1) (2) (3) Wherein, A is the water area, P is the wet perimeter, and R is the hydraulic radius. is the Froude number, is the flow rate, is the water depth, is the gravitational acceleration. is the Reynolds number, is the kinematic viscosity coefficient. After the deep coupling of physical laws and data features is completed, the sliding mean features of, for example, 6, 12 and 24 periods, the spatial lag features of 1-4 periods, and other spatial features are extracted to construct a comprehensive feature matrix containing time, space, physics and statistics as shown in formula (4): Figure 2 The comprehensive feature matrix facilitates the model to realize sediment deposition prediction based on the complex physical process of sediment transport. Wherein, Figure 2 is a distribution diagram of the comprehensive feature matrix in a three-dimensional feature space in the multi-source data fusion and feature extraction link, reflecting the distribution structure and discrimination ability of the four types of features of time, space, physics and statistics in the irrigation area sediment deposition prediction method provided by the application after deep fusion. Specifically, Figure 2Three types of features are selected from the comprehensive feature system, which correspond to the three feature dimensions of physics, statistics and space. The X-axis selects the flow as the representative of the physical feature, reflecting the core variable of the hydrodynamic process, the Y-axis selects the 6-period sliding mean of the flow as the representative of the statistical feature, embodying the time sequence smoothing and trend information of the hydrological process, and the Z-axis selects the distance between sections as the representative of the spatial feature, reflecting the heterogeneity of the spatial structure and section distribution of the river channel. Figure 2 Each point in the figure represents a spatio-temporally aligned observation sample, and the color reflects the risk level and other label features. These samples are derived from multi-source heterogeneous data such as historical hydrology, real-time monitoring, remote sensing observation, etc., and after spatio-temporal alignment, feature extraction and standardization processing, a high-quality feature matrix is formed.

[0025] The method for predicting sediment deposition in an irrigation area provided by the present application considers the limitations of a single data source, collects historical hydrological data, real-time monitoring data and remote sensing observation data of the irrigation area to be predicted as multi-source data input into the model for feature extraction, obtains irrigation area physical features, irrigation area statistical features, irrigation area spatial features and irrigation area time features that can fully reflect the overall sediment transport process of the irrigation area, and performs sediment deposition prediction after fusing the four types of features, thereby improving the accuracy of the prediction of sediment deposition in the irrigation area.

[0026] In addition, considering the problem of uneven spatio-temporal distribution among multi-source data and direct prediction deviation being too large, the method for predicting sediment deposition in an irrigation area provided by the present application resamples the historical hydrological data and real-time monitoring data according to a unified time axis, and fills the remote sensing observation data into the unified time axis, thereby achieving spatio-temporal alignment among multi-source data and avoiding the influence of uneven spatio-temporal distribution of data on the prediction of sediment deposition.

[0027] Based on the above embodiments, as shown in Figure 3 The method for predicting sediment deposition in an irrigation area provided by the present application comprises the following steps: Step 111, resampling the historical hydrological data according to a unified time axis in the monitoring section of the irrigation area to be predicted; and resampling the real-time monitoring data according to a unified time axis.

[0028] Step 112, when the historical hydrological data lacks observation values within the time reference of the unified time axis, filling the historical hydrological data observation values of the last time reference to obtain resampled historical hydrological data.

[0029] Step 113, when the real-time monitoring data lacks observation values within the time reference of the unified time axis, filling the real-time monitoring data observation values of the last time reference to obtain resampled real-time monitoring data.

[0030] And, as shown in Figure 4As shown, the application provides a kind of irrigation silt deposition prediction method, and step 11 further comprises: Step 114, when the monitoring section lacks spatial features within the time reference of the unified time axis, fill in based on the spatial feature values of the monitoring section at the last time reference.

[0031] Step 115, when the monitoring section lacks spatial feature values at the last time reference, fill in based on the spatial feature values of the monitoring section at the next time reference.

[0032] Step 116, when the monitoring section lacks spatial feature values at the next time reference, inverse distance weighting is performed on the spatial feature values of the adjacent monitoring section to obtain the remote sensing observation data with completed spatial features.

[0033] Specifically, considering that the sampling frequency and time stamp of the actual historical hydrological data, real-time monitoring data and remote sensing observation data of the to-be-predicted irrigation area are often inconsistent, in order to realize the fusion of multi-source data, a unified time reference needs to be set first, for example, the time reference is set to 1 hour, and a standard time axis covering the whole time period is generated according to this, and all data are aligned on this standard time axis, to ensure the time sequence consistency of subsequent feature extraction and modeling.

[0034] For example, in each monitoring section of the to-be-predicted irrigation area, the historical hydrological data and real-time monitoring data are resampled according to the unified time axis. The resampling strategy of the historical hydrological data and real-time monitoring data is a forward filling strategy, when the historical hydrological data or real-time monitoring data at a certain time lacks observation values, the historical hydrological data or real-time monitoring data at the last time is filled in as effective observation values at this time, as shown in formula (4):

[0035] (4) Wherein, represents the observation value of the th section at time , if missing, fill in the value of the last time. This forward filling method guarantees the continuity of the data, which is suitable for the engineering practice of hydrological sequence. For remote sensing observation data, considering that remote sensing data is usually sampled sparsely and irregularly, for each monitoring section, latitude, longitude, river width, curvature and other spatial features are extracted from the remote sensing observation data, and these spatial features are filled in at all time points, so that complete spatial description can be obtained at each time.

[0036] Wherein, the remote sensing observation data also has the defect of spatial feature missing, which can be aligned in space and completed in space by the following method: First, a standard spatial grid is constructed based on the actual monitoring sections deployed in the irrigation area, and the positions are calibrated using remote sensing data to ensure that each monitoring section has accurate and unique spatial coordinates or latitude and longitude, thereby ensuring the accuracy of the spatial distribution of the sections. Then, the geodesic distance formula is used to calculate the spatial distance between each monitoring section and its nearest neighbor based on the geographic coordinates of each section. The geodesic distance formula is shown in formula (5):

[0037] (5) in, cross-section and cross-section The spherical distance between them For the Earth's radius, and These represent the latitude and longitude of the cross-section, respectively. Spherical distance not only reflects the density of cross-section distribution but also forms the basis for subsequent calculations of spatial characteristics such as spatial lag and spatial correlation.

[0038] For sections lacking spatial attributes, three methods can be used to fill them: forward filling, backward filling, or spatial interpolation. For example, the forward filling method is preferred, which fills the section with the valid observations at the previous time. This method is suitable for situations where the spatial attributes change little over time and the section itself has historical observations. If there are still missing attributes after forward filling, the backward filling method is used, which fills the section with the valid observations at the next time. This method is suitable for scenarios such as the initial stage of data acquisition. If a section has no valid observations at all times, the spatial interpolation method is used, which references the observations of adjacent sections in space and performs weighted average filling by inverse distance weighting. The weighted average filling formula is shown in formula (6):

[0039] (6) in, To provide the spatial characteristics of the section to be completed, Given the known characteristic values ​​of adjacent cross sections, The spatial distance from the section to be completed, The weighting index is 2 in this embodiment. For example, if a section is missing the width of the river channel, the width of the river channel of the adjacent sections can be weighted and averaged according to the spatial distance between the section and the adjacent sections. The closer the section is, the greater the weight, thereby achieving reasonable completion of the spatial attributes.

[0040] In addition, considering the upstream flow influence, the embodiment can also automatically identify the upstream section in combination with the section spatial distribution and calculate the weighted influence of the upstream section on the current section flow. The weight is usually inversely proportional to the distance between the sections, and the closer the distance, the greater the influence. Meanwhile, for each section, the flow, sediment concentration, water depth and other historical observation values of the upstream 1-4 stages are extracted as spatial lag features. These features can capture the spatial memory effect of sediment transport and help improve the model's expression ability for complex physical processes.

[0041] When the three types of data are resampled and the spatial features are completed, all the data are merged with "section ID + timestamp" as the primary key. For the same section at the same time, the real-time monitoring data are preferred to complete the missing of the historical data, and the spatial features are completed by the remote sensing data. In this way, a multi-source data set with spatio-temporal alignment and complete features is finally formed. Moreover, when the data are aligned, the data can also be physically consistent by using the continuity principle of water flow, and the verification formula is shown in formula (7):

[0042] (7) Wherein, and are the flow of the upstream and downstream sections at time , respectively, is the water volume change between the sections, is the time interval. If the difference is too large, the data is abnormal and needs to be further corrected.

[0043] On the basis of the above-mentioned embodiment, a silt deposition prediction method provided by the application, the silt deposition prediction result is obtained by processing the physical characteristics of the irrigation area, the statistical characteristics of the irrigation area, the spatial characteristics of the irrigation area and the time characteristics of the irrigation area through the silt transport prediction model, wherein the silt transport prediction model is composed of an input layer, a position coding layer, a Transformer encoder, a Transformer decoder, a residual connection layer and an output layer connected in turn. The position coding layer is used for time sequence enhancement of the physical characteristics of the irrigation area, the statistical characteristics of the irrigation area, the spatial characteristics of the irrigation area and the time characteristics of the irrigation area, and the time sequence feature coding of the Transformer encoder. The Transformer encoder is a multi-layer stacked structure, and each layer of the Transformer encoder includes a multi-head self-attention mechanism module and a feedforward neural network.

[0044] And, the application provides a silt accumulation prediction method for an irrigation area, wherein the silt transport prediction model further comprises a physical constraint layer, an input end of the physical constraint layer is connected with an output end of the Transformer decoder, an output end of the physical constraint layer is connected with an input end of the residual connection layer, wherein the physical constraint layer integrates the Manning coefficient and the silt settling coefficient, and is used for calculating the silt transport prediction model through a physical risk calculation formula.

[0045] Specifically, the application provides a silt accumulation prediction method for an irrigation area, wherein a multi-level model structure optimization scheme is designed based on a traditional Transformer architecture and specific requirements of silt accumulation prediction in the irrigation area, so as to realize multi-target joint prediction of silt accumulation in the irrigation area. Figure 5 As shown in FIG. 52, when the 52-dimensional physical-data fusion features are input into the model, the fusion features are first processed through a feature projection layer and then encoded by a Transformer encoder. The Transformer encoder adopts 6-layer Transformer encoder layers, each layer containing a multi-head self-attention mechanism with, for example, 8 attention heads and a feedforward neural network, and the model's ability to capture time sequence features is enhanced through position encoding and cyclic position encoding. For example, the model configuration parameters include a hidden dimension of 256, 8 attention heads, 6 layers, a dropout rate of 0.1, and a total parameter amount of 54.68 million, which provides sufficient expression ability for complex spatiotemporal modeling. Then, the model is decoded by a Transformer decoder also with 6-layer Transformer decoder layers, each layer containing a multi-head self-attention mechanism with, for example, 8 attention heads, and the input features and the output layer are directly connected through a residual connection mechanism to ensure that the model can retain the original physical feature information and improve the prediction accuracy and generalization ability. Finally, the multi-target prediction results of silt concentration, flow velocity and flow rate are determined based on the decoding results.

[0046] The feature projection layer is a key component in the model structure, and is mainly used to map the input high-dimensional, multi-source fusion features to a unified feature space suitable for processing by the Transformer model. On the one hand, since the initial input is 52-dimensional physical-data fusion features, these features are diverse in origin, physical dimension, numerical range, and statistical distribution. The feature projection layer can perform linear transformation through a fully connected layer or a linear layer to map the original features to a required 256-dimensional hidden space in the model, achieving uniformity of feature dimensions and information compression. On the other hand, through feature projection, the model can automatically learn the optimal combination and weighting method between different features, improving the expression ability of the features and providing more abundant input information for the subsequent Transformer encoder to capture the spatio-temporal dependency. In addition, the feature projection layer can also filter out noise and redundant information in the input features to some extent, allowing the model to focus on the feature subspace that is most useful for predicting sediment deposition, thereby improving the generalization ability and prediction accuracy of the model.

[0047] For example, in the model structure of the present embodiment, the feature projection layer has functions such as linear transformation, layer normalization, Gaussian error linear unit (GELU) activation, and Dropout regularization. First, the linear transformation technology of the fully connected layer is used to map the original 52-dimensional multi-source fusion features to a high-dimensional hidden space, achieving uniformity of feature dimensions and information compression, and allowing features of different origins and dimensions to be effectively expressed and combined in the same space. Second, the layer normalization technology is applied to normalize the hidden features of each sample. This helps to improve the stability of model training, accelerate the convergence speed, and alleviate the problem of gradient vanishing or explosion in deep networks. Then, the GELU activation function is introduced to enhance the expression ability of the model for complex feature relationships, making the network have stronger fitting and generalization ability. Finally, the integrated Dropout regularization technology is used to randomly discard some neurons, effectively improving the robustness and generalization ability of the model.

[0048] And the model adopts an adaptive learning rate scheduling strategy, combines a single cycle learning rate scheduler and a learning rate decay scheduler, and realizes fine optimization of the sediment prediction task. The training configuration includes parameters such as a batch size of 64, a learning rate of 0.0001, a weight decay of 0.0001, a gradient clipping of 1.0, and an early stop patience value of 10. The model supports the combination of relative MSE loss and Huber loss, and the model performance is monitored through evaluation indicators such as a relative error threshold of 0.2, an absolute error threshold of 0.5, and an NSE threshold of 0.7. Moreover, the loss function is a multi-objective loss function during the model training stage, and different loss weights are set for the three prediction targets of sediment concentration, flow velocity and flow rate, for example, 0.6, 0.25 and 0.15 respectively, which effectively balances the accuracy requirements of different prediction targets. When the loss weight is determined, the model can store the loss weight for subsequent model application, and generate a multi-objective joint prediction result based on the loss weight.

[0049] And on the basis of the Transformer model, the application also sets a physical constraint layer between the Transformer decoder and the residual connection layer. The physical constraint layer integrates key physical parameters such as a Manning coefficient set to 0.025 and a sediment settling velocity set to 0.12 m / s, and performs physical consistency constraint on the model output through a physical risk calculation formula. The physical risk calculation is shown in formula (8):

[0050] (8) wherein, is a physical risk calculation result, is a current flow, is a critical flow threshold, is a current sediment concentration, is a critical sediment concentration threshold, is a current flow velocity, is a critical flow velocity threshold, , , is a physical parameter weight coefficient.

[0051] Specifically, addressing the complex data requirements of the 52-dimensional feature system of irrigation districts, this invention integrates historical hydrological, real-time monitoring, and remote sensing observation data through a multi-source data fusion module, solving the prediction bias problem caused by the single data source in traditional methods. The system automatically extracts physical features such as flow rate, water depth, sediment concentration, flow velocity, channel width, and tortuosity, as well as statistical features such as moving average, spatial lag, and spatiotemporal correlation, improving the prediction accuracy for heterogeneous areas of the irrigation district to over 85%. Considering the spatiotemporal distribution characteristics of multiple sections and multiple time points in the irrigation district, this invention uses a spatiotemporal alignment processor to align the data between sections and time points, solving the prediction errors caused by uneven spatiotemporal distribution in traditional methods. The system automatically calculates irrigation district-specific parameters such as upstream flow influence and spatial lag characteristics, addressing the problem of general models neglecting the spatial structure of the irrigation district. Furthermore, the introduction of a physical constraint layer and residual connection mechanism into the model deeply integrates hydraulic physics laws with deep learning models, further improving the accuracy of sediment deposition prediction.

[0052] Based on the above implementation methods, such as Figure 6 As shown, in the method for predicting sediment deposition in irrigation areas provided by this invention, after step 13, the method further includes: Step 14: Analyze the causes of sediment deposition by using interpretability analysis tools and pre-set causal relationships for sediment deposition, including sediment flow rate, sediment concentration, and sediment velocity, and obtain the results of the sediment deposition cause analysis.

[0053] And, as Figure 7 As shown, in the method for predicting sediment deposition in irrigation areas provided by this invention, after step 13, the method further includes: Step 15: Based on sediment flow rate, sediment concentration and sediment flow velocity, conduct a sediment deposition risk assessment of the irrigation area to be predicted, and obtain the risk classification and risk level of the irrigation area to be predicted.

[0054] Step 16: Generate gate scheduling strategy for the irrigation area to be predicted based on the risk classification and risk level of the irrigation area to be predicted.

[0055] Specifically, the siltation prediction method for irrigation areas provided by this invention also integrates the SHAP interpretability analysis tool based on the Transformer architecture, realizing transparent explanation of the model decision-making process. The SHAP value calculation formula is shown in formula (9):

[0056] (9) in, For the first The SHAP value of each feature, This represents the model prediction value when all features are included. To remove the first The model's predicted value for each feature. For example... Figure 8As shown, in one aspect, the model quantifies the contribution of each feature to the prediction result through SHAP explainability analysis tool, reveals the key influencing factors of sediment concentration, flow velocity and flow rate prediction through feature importance analysis, and builds a complete causal path from upstream flow to hydraulic parameters to sediment transport to siltation risk through causal chain analysis, supporting the traceability analysis of abnormal prediction results. Among them, the causal path model contains 15 kinds of causal relationship rules, covering multiple dimensions such as hydrological characteristics, river channel morphology, hydraulic characteristics and spatiotemporal characteristics, and achieves high precision of Nash efficiency coefficient NSE>0.97 in sediment concentration prediction, while having complete explainability, providing scientific and reliable theoretical support for engineering decision-making.

[0057] On the other hand, the irrigation sediment siltation prediction method provided by the present application can also perform risk assessment on the model prediction risk and physical risk calculation results corresponding to the model prediction results, and the comprehensive risk index calculation is as shown in formula (10): (10) Among them, is the AI prediction risk, and are the fusion weight coefficients of physical risk and AI prediction risk. After determining the comprehensive risk index, the flow adjustment coefficient can also be automatically generated based on the risk level corresponding to the comprehensive risk index, for example, 1.8 times for critical risk, 1.5 times for high risk, and 1.15-1.3 times for medium risk, and fine gate scheduling strategies such as flow adjustment, operation delay and staged desilting are generated.

[0058] In addition, the irrigation sediment siltation prediction method provided by the present application can also realize real-time interaction with external platforms through a digital twin interface, support automatic issuance and execution monitoring of control instructions, and provide a complete technical solution for intelligent and fine management of irrigation water and sediment regulation.

[0059] The irrigation sediment siltation prediction method provided by the present application analyzes the causes of irrigation sediment siltation through SHAP explainability analysis and 15 kinds of causal relationship rules, providing a scientific decision-making basis for irrigation managers. The causal chain visualization report and feature importance analysis results generated by the system improve the transparency of management decisions by 40%, meeting the actual needs of fine management of irrigation areas. Specifically, the irrigation sediment siltation prediction method provided by the present application has built-in SHAP explainability analysis tool and causal relationship chain analysis module. Global and local feature importance analysis can be performed on the model prediction results, automatically generating feature dependency relationship, spatiotemporal influence heat map and causal chain visualization report, and supporting traceability analysis of key features and abnormal samples, improving the transparency of the model and the credibility of the decision.

[0060] And, the irrigation area silt accumulation prediction method provided by the present application realizes real-time interaction with external platforms through a digital twin interface, supports irrigation area-specific scheduling strategies such as flow adjustment, operation delay, and staged sediment discharge, and improves the regulation efficiency by 35%, thereby adapting to the periodic management needs of irrigation area water and sediment scheduling. Specifically, the irrigation area silt accumulation prediction method provided by the present application is equipped with a lightweight visualization engine, which supports multiple visualization forms such as risk heat maps, feature importance bar charts, SHAP force diagrams, dependency diagrams, and animation demonstrations. Static and dynamic prediction result displays can be automatically generated, which facilitates users to intuitively understand the model output and regulation effect. All visualization results can be exported as pictures or HTML animations, supporting integration with external systems.

[0061] In addition, the irrigation area silt accumulation prediction method provided by the present application innovatively deeply integrates physical rules and AI prediction results, and constructs an intelligent regulation strategy generation mechanism based on risk levels, thereby solving the core defects of insufficient physical-data fusion and limited system adaptability in the prior art. Specifically, the irrigation area silt accumulation prediction method provided by the present application integrates a risk assessment and intelligent regulation module based on physical-model fusion. Through model prediction results and physical rule calculations, a risk index for each section and each time is comprehensively generated. The regulation strategy module automatically generates optimal gate scheduling instructions according to the risk level, spatial partitioning, and gate distribution, including flow adjustment, operation delay, and staged sediment discharge.

[0062] Embodiment 2 Based on embodiment 1, as shown in Figure 9 The present application also provides a specific irrigation area silt accumulation prediction example: First, the data acquisition domain preprocessing stage collects historical hydrological data, real-time monitoring data, and remote sensing observation data of the irrigation area to be predicted, and performs standardization preprocessing between multiple sources of data. Subsequently, the standardized preprocessed multi-source data is subjected to multi-source data fusion and spatio-temporal alignment, and 52-dimensional physical-data fusion features are extracted and the feature set is standardized to obtain 52-dimensional physical-data fusion features.

[0063] Then, the 52-dimensional physical-data fusion features are input into the pre-trained deep learning model, and the 52-dimensional physical-data fusion features are converted into vector representations rich in context information based on the Transformer encoder, capturing global dependencies within the sequence, and capturing context awareness by the multi-head attention mechanism. Then, the Transformer decoder gradually generates the target sequence based on the output results of the Transformer encoder and performs silt accumulation prediction. After the model output is physically consistent by the deep learning model, the final multi-objective prediction result is output.

[0064] Then, SHAP value calculation is performed by the SHAP explainability analysis tool, feature importance analysis and causal chain analysis are performed based on the SHAP value calculation result, risk assessment and risk level classification are performed based on the physical risk result calculated by the physical constraint layer and the risk result determined by the model, and a gate scheduling strategy is generated based on the risk level classification result, and finally the gate scheduling strategy is output through a digital twin interface.

[0065] In addition, the irrigation district sediment deposition prediction method provided by the present application can also output and display the prediction results through an intelligent display platform integrating various dynamic visualization functions such as risk thermodynamic maps, feature importance maps, and prediction animations. For example, spatial partitioning is achieved through the Venn diagram algorithm, and risk thresholds are automatically generated for low-risk R < 0.4, medium-risk 0.4 ≤ R < 0.7, high-risk 0.7 ≤ R < 0.9, and critical-risk R ≥ 0.9. In addition, the intelligent display platform supports real-time generation of HTML animations to show the spatio-temporal evolution of sediment deposition risk, and integrates SHAP force diagrams, dependency diagrams, and other explainability visualization components to support result export and external system integration.

[0066] In summary, the irrigation district sediment deposition prediction method provided by the present application integrates multiple key technologies such as multi-source data fusion, feature engineering, spatio-temporal alignment, deep learning modeling, intelligent decision-making, explainability analysis, and efficient visualization, achieving high-precision prediction and intelligent control of sediment deposition risk in irrigation districts. The system can automatically load, clean, extract features, and standardize processing of multi-source heterogeneous data such as historical hydrology, real-time monitoring, and remote sensing, ensuring data integrity and consistency. The deep learning model based on the Transformer structure jointly models multi-dimensional spatio-temporal features, effectively capturing the complex spatio-temporal evolution of sediment deposition and significantly improving the accuracy and stability of risk prediction.

[0067] The model of the system has multi-target prediction capability and can simultaneously predict key indicators such as sediment concentration or deposition risk, flow rate, and flow volume at each section and time in the irrigation district. According to actual training and test results, the NSE of the model in the sediment concentration, flow rate, and flow volume prediction tasks is greater than 0.97, showing extremely high goodness of fit and prediction reliability. At the same time, the accuracy rate of the sediment concentration prediction on the validation set is stable at 0.85 or above, the flow rate prediction accuracy rate is about 0.98, and the flow volume prediction accuracy rate is stable at 0.85 or above, with good overall loss value convergence and excellent generalization ability. The above results show that the system can provide high-precision, real-time multi-target prediction results for irrigation district management, meeting the needs of engineering applications.

[0068] Combined with physical rules and model outputs, the system can automatically generate gate control strategies for different risk levels, realizing intelligent and refined management of water and sediment dispatch. The integrated explainable analysis tool can output feature importance, dependency and causal chain analysis results, improving the transparency and credibility of model decision-making. The supporting visualization module can intuitively display risk distribution, control effect and feature analysis results, facilitating managers to quickly grasp the operation state of the irrigation district. The overall scheme realizes full-process automation from data processing, model training, risk prediction, decision generation to visualization display, has high modularity and scalability, can effectively improve the early warning ability and control efficiency of the sediment accumulation risk of the irrigation district, and guarantee the safety and water resource utilization efficiency of the operation of the irrigation district, and has good engineering application prospect.

[0069] Embodiment 3 The application further provides an irrigation district sediment accumulation prediction system, comprising: An irrigation district data acquisition module is configured to acquire historical hydrological data, real-time monitoring data and remote sensing observation data of an irrigation district to be predicted, resample the historical hydrological data and the real-time monitoring data according to a unified time axis of the irrigation district to be predicted, complete the historical hydrological data or the real-time monitoring data that are missing in the unified time axis by forward filling, fill the remote sensing observation data into the unified time axis, and fill the spatial features that are missing in the remote sensing observation data based on the spatial feature values of adjacent time references on the unified time axis to obtain multi-source data aligned in time and space, wherein the unified time axis covers the whole time period of the irrigation district to be predicted.

[0070] A multi-source feature extraction module is configured to extract features from the multi-source data aligned in time and space to obtain irrigation district engineering morphological features, irrigation district hydrological index statistical features, irrigation district spatial grid features and irrigation district time series operation features.

[0071] A sediment accumulation prediction module is configured to combine the irrigation district engineering morphological features, the irrigation district hydrological index statistical features, the irrigation district spatial grid features and the irrigation district time series operation features, calculate the correlation weights between the features based on a multi-head attention mechanism, obtain an attention feature matrix, and determine the sediment flow, the sediment concentration and the sediment flow rate of the irrigation district to be predicted according to the attention feature matrix.

[0072] The application further provides a computer device comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to realize the steps in an embodiment of the irrigation district sediment accumulation prediction method. For specific implementation methods, refer to the method embodiments, which will not be described here.

[0073] Further, the present application also provides a non-transitory computer readable storage medium comprising instructions. For example, a memory comprising instructions that are executable by a processor of a computer device to perform the above method. For example, the non-transitory computer readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disc, and an optical data storage device, etc. The computer program is executed by the processor to implement the steps in the embodiment of the method for predicting sediment deposition in an irrigation district. For details, refer to the method embodiments, which will not be described here.

[0074] Those skilled in the art should understand that the embodiments of the present application can provide a method, a system or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0075] The present application is described with reference to the flowcharts and / or block diagrams of the method, device (system) and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device implemented in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 a device configured to perform the function specified in the flow or flows and / or block or blocks.

[0076] These computer program instructions can also be stored in a computer readable memory that can direct the computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce a manufactured product including instruction devices that implement the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 a device configured to perform the function specified in the flow or flows and / or block or blocks.

[0077] These computer program instructions can also be loaded into a computer or other programmable data processing device to cause a series of operation steps to be performed on the computer or other programmable data processing device to produce a computer implemented process, so that the instructions executed on the computer or other programmable data processing device provide a process for implementing the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1steps of the functions specified in the block or blocks.

[0078] It should be noted that the above detailed description and accompanying examples are indicative, for those skilled in the art, of the full and enabling disclosure of the application created. No limitation, equivalent substitution or change, in form and details, of the herein disclosed application can be accepted under any circumstance by anyone without the approval of the patentee. Any reference signs in the claims should not be construed as limiting the claims. The simple changes or equivalent substitutions or changes, which can be obviously obtained by those skilled in the art within the technical scope disclosed by the application, shall fall within the protection scope of the application.

Claims

1. A method for predicting sediment deposition in irrigation areas, characterized in that, include: Collect historical hydrological data, real-time monitoring data, and remote sensing observation data of the irrigation area to be predicted; The historical hydrological data and real-time monitoring data are resampled according to the unified timeline of the irrigation area to be predicted. The missing historical hydrological data or real-time monitoring data in the unified timeline are filled in by forward filling. The remote sensing observation data is filled into the unified timeline. The missing spatial features in the remote sensing observation data are filled in based on the spatial feature values ​​of the adjacent time reference on the unified timeline to obtain spatiotemporally aligned multi-source data. The unified timeline covers the entire time period of the irrigation area to be predicted. Feature extraction was performed on spatiotemporally aligned multi-source data to obtain the engineering morphology features of irrigation districts, the statistical features of hydrological indicators of irrigation districts, the spatial grid features of irrigation districts, and the temporal operational features of irrigation districts. The irrigation district engineering morphology features, irrigation district hydrological index statistical features, irrigation district spatial grid features, and irrigation district temporal operation features are combined and the correlation weights between features are calculated based on a multi-head attention mechanism to obtain an attention feature matrix. The sediment flow, sediment concentration, and sediment velocity of the irrigation district to be predicted are determined based on the attention feature matrix.

2. The method for predicting sediment deposition in irrigation areas according to claim 1, characterized in that, The historical hydrological data and real-time monitoring data are resampled according to a unified timeline of the irrigation area to be predicted. Missing historical hydrological data or real-time monitoring data in the unified timeline are filled in by forward imputation, including: In the monitoring section of the irrigation area to be predicted, the historical hydrological data are resampled according to the unified time axis; the real-time monitoring data are also resampled according to the unified time axis. When the historical hydrological data lacks observations within the time reference of the unified time axis, it is filled with historical hydrological data observations based on the previous time reference to obtain resampled historical hydrological data. When the real-time monitoring data lacks observations within the time reference of the unified time axis, it is filled with real-time monitoring data observations from the previous time reference to obtain resampled real-time monitoring data.

3. The method for predicting sediment deposition in irrigation areas according to claim 1, characterized in that, Filling in the missing spatial features in the remote sensing data based on spatial feature values ​​of adjacent time bases on the unified time axis includes: When the monitoring section lacks spatial features within the time reference of the unified time axis, it is filled based on the spatial feature values ​​of the monitoring section in the previous time reference. When the monitoring section lacks spatial feature values ​​in the previous time reference, it is filled based on the spatial feature values ​​of the monitoring section in the next time reference. When the monitoring section lacks spatial feature values ​​at the next time reference, the spatial feature values ​​of adjacent monitoring sections are inversely distance-weighted to obtain remote sensing observation data with complete spatial features.

4. The method for predicting sediment deposition in irrigation areas according to claim 1, characterized in that, The sediment transport prediction model processes the physical characteristics, statistical characteristics, spatial characteristics, and temporal characteristics of the irrigation area to obtain the sediment deposition prediction results. The sediment transport prediction model consists of an input layer, a position encoding layer, a Transformer encoder, a Transformer decoder, a residual connection layer, and an output layer connected in sequence.

5. The method for predicting sediment deposition in irrigation areas according to claim 4, characterized in that, The sediment transport prediction model also includes a physical constraint layer. The input of the physical constraint layer is connected to the output of the Transformer decoder, and the output of the physical constraint layer is connected to the input of the residual connection layer. The physical constraint layer integrates the Manning coefficient and the sediment settling coefficient, and is used to calculate the sediment transport prediction model using a physical risk calculation formula.

6. The method for predicting sediment deposition in irrigation areas according to claim 1, characterized in that, After determining the sediment flow rate, sediment concentration, and sediment velocity of the irrigation area to be predicted, the method further includes: The sediment flow rate, sediment concentration, and sediment velocity were analyzed using interpretability analysis tools and pre-defined causal relationships to determine the causes of sediment deposition, and the results of the sediment deposition cause analysis were obtained.

7. The method for predicting sediment deposition in irrigation areas according to claim 1, characterized in that, After determining the sediment flow rate, sediment concentration, and sediment velocity of the irrigation area to be predicted, the method further includes: Based on the sediment flow rate, sediment concentration and sediment flow velocity, a sediment deposition risk assessment is conducted for the irrigation area to be predicted, and the risk classification and risk level of the irrigation area to be predicted are obtained. A gate scheduling strategy for the irrigation area to be predicted is generated based on the risk classification and risk level of the irrigation area to be predicted.

8. A sediment deposition prediction system for irrigation areas, characterized in that, include: The irrigation district data acquisition module is used to collect historical hydrological data, real-time monitoring data, and remote sensing observation data of the irrigation district to be predicted. The historical hydrological data and real-time monitoring data are resampled according to the unified timeline of the irrigation area to be predicted. The missing historical hydrological data or real-time monitoring data in the unified timeline are filled in by forward filling. The remote sensing observation data is filled into the unified timeline. The missing spatial features in the remote sensing observation data are filled in based on the spatial feature values ​​of the adjacent time reference on the unified timeline to obtain spatiotemporally aligned multi-source data. The unified timeline covers the entire time period of the irrigation area to be predicted. The multi-source feature extraction module is used to extract features from spatiotemporally aligned multi-source data to obtain irrigation area engineering morphology features, irrigation area hydrological index statistical features, irrigation area spatial grid features, and irrigation area temporal operation features. The sedimentation prediction module is used to merge the engineering morphology features, hydrological index statistical features, spatial grid features, and temporal operation features of the irrigation area, and calculate the correlation weights between the features based on a multi-head attention mechanism to obtain an attention feature matrix. Based on the attention feature matrix, the sediment flow, sediment concentration, and sediment velocity of the irrigation area to be predicted are determined.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the irrigation area siltation prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by the processor, it is able to execute the steps of the siltation prediction method for irrigation areas as described in any one of claims 1 to 7.

Images