Water conservancy project system real-time scheduling method based on neural network technology

Abstract

The invention relates to the technical field of flood prevention management, in particular to a hydraulic engineering system real-time scheduling method based on a neural network technology, and the method comprises the steps: obtaining the real-time data of a hydraulic engineering system at a fixed time interval; performing image recognition on the satellite cloud picture by adopting a CNN model to obtain a rain cloud region map and a rain cloud region edge map; edge roughness coefficients, internal density coefficients and rain cloud thickness coefficients of the rain cloud area map and the rain cloud area edge map are calculated, so that rainfall coefficients are obtained; analyzing the rain cloud area map at continuous time intervals by using an optical flow method, obtaining the moving speed and direction of the rain cloud and the shortest distance between the rain cloud and the water conservancy project area, and obtaining a rain cloud influence coefficient; predicting the future water level of the reservoir by using a deep learning model, and calculating a vulnerability coefficient according to the flood control data; the flood control coefficient is calculated according to the rainfall coefficient, the rain cloud influence coefficient and the vulnerability coefficient, and the flood control regulation and control model is constructed, so that the flexibility of real-time scheduling of the hydraulic engineering system is effectively improved.

Description

Technical Field

[0001] The invention relates to the technical field of flood control management, and in particular to a real-time dispatching method for a water conservancy project system based on a neural network technology. Background Art

[0002] Traditional water conservancy project scheduling methods usually rely on experience and historical data and are difficult to adapt to real-time changing environments. This leads to insufficient response during floods, especially in extreme weather conditions, and the inability to respond quickly to sudden rainfall, which affects the reservoir's water storage and flood discharge decisions, and cannot effectively ensure the safety of the reservoir and the stability of water supply. In order to improve the efficiency and flexibility of water conservancy project scheduling, more and more research has begun to focus on real-time scheduling methods, which improve the responsiveness of scheduling. However, existing methods often lack a comprehensive analysis of rain clouds, terrain factors, and social factors, and lack the use of this information for effective reservoir management, which limits the dynamics and accuracy of scheduling strategies, especially in the face of complex meteorological changes and multiple scheduling constraints. Flexibility needs to be further improved.

[0003] Therefore, a real-time dispatching method for water conservancy engineering system based on neural network technology is proposed. Summary of the invention

[0004] The purpose of the present invention is to provide a real-time scheduling method for a water conservancy project system based on a neural network technology, which obtains real-time data of the water conservancy project system at fixed time intervals; uses a CNN model to perform image recognition on a satellite cloud image to obtain a rain cloud area map and a rain cloud area edge map; calculates the edge roughness coefficient, internal density coefficient and rain cloud thickness coefficient of the rain cloud area map and the rain cloud area edge map, thereby obtaining a rainfall coefficient; uses an optical flow method to analyze the rain cloud area map at continuous time intervals, obtains the moving speed and direction of the rain cloud and the shortest distance between the rain cloud and the water conservancy project area, and obtains the rain cloud influence coefficient; uses a deep learning model to predict the future water level of the reservoir, and calculates the vulnerability coefficient based on flood control data; calculates the flood control coefficient based on the rainfall coefficient, the rain cloud influence coefficient and the vulnerability coefficient, and constructs a control model, thereby effectively improving the flexibility of real-time scheduling of the water conservancy project system.

[0005] To achieve the above object, the present invention provides the following technical solutions: A real-time dispatching method for a water conservancy project system based on neural network technology, comprising: Acquire real-time data of a water conservancy project system at fixed time intervals, the real-time data including satellite cloud images, rain cloud vertical profile data and flood control data, and pre-process the satellite cloud images to obtain a first image data set; The first image data set is subjected to image recognition using a CNN model, wherein the CNN model is used to extract spatial features of rain clouds, and obtain a rain cloud region map and a rain cloud region edge map; the image recognition process is as follows: feature extraction is performed according to the first image data set to obtain low-level image features; a first branch network and a second branch network are formed according to the low-level image features, wherein the first branch network is used to segment the rain cloud region and output the rain cloud region map, and the second branch network is used to extract the edge of the rain cloud region and output the rain cloud region edge map; An edge roughness coefficient and an internal density coefficient are calculated based on the rain cloud area map and the rain cloud area edge map; a rain cloud thickness coefficient is calculated based on the rain cloud vertical profile data, and a rainfall coefficient is calculated based on the edge roughness coefficient, the internal density coefficient and the rain cloud thickness coefficient; The rain cloud area map at continuous time intervals is calculated using the optical flow method to obtain the rain cloud moving speed and the rain cloud moving direction, and the shortest distance between the central rain cloud in the rain cloud area map and the water conservancy project area is calculated; the rain cloud influence coefficient is calculated according to the rain cloud moving speed, the rain cloud moving direction and the shortest distance; Predicting the future water level of the reservoir using a deep learning model based on the flood control data, and calculating based on the flood control data and the future water level of the reservoir to obtain a vulnerability coefficient; A flood control coefficient is calculated based on the rainfall coefficient, the rain cloud influence coefficient and the vulnerability coefficient, and a flood control model is established based on the flood control coefficient. The flood control model is used to perform real-time control on the water conservancy project system.

[0006] Furthermore, the flood control data include: flood defense historical data, reservoir historical data, reservoir real-time data, topographic and geographical data, building density, population density and hydrological facilities.

[0007] Furthermore, the preprocessing includes: Performing color space conversion on the satellite cloud image to obtain a first cloud image; Correcting the illumination and radiation of the first cloud image to obtain a second cloud image; Determine a downsampling ratio according to task requirements, and use bilinear interpolation to downsample the second cloud image according to the downsampling ratio to obtain a third cloud image; The third cloud image is subjected to a Gaussian filter to remove noise, thereby obtaining the first image data set.

[0008] Furthermore, the training steps of the CNN model include: Step S1: Acquire historical satellite cloud images and perform the preprocessing to obtain a historical cloud image data set; Step S2: labeling the historical cloud image dataset with image labels to obtain a second image dataset, and inputting the second image dataset into the CNN model; wherein the image labels include rain cloud areas, non-rain cloud areas, edge points, and non-edge points; Step S3: using a convolution layer to perform a convolution operation on the second image data set, and output a first feature map; Step S4: using a ReLU activation function and a pooling layer to perform a pooling operation on the first feature map, and output a second feature map; Step S5: using the first branch network to perform a segmentation operation on the second feature map, and outputting the rain cloud area map; Step S6: using the second branch network to perform edge feature detection on the second feature map, and outputting the rain cloud area edge map; Step S7: using the cross entropy loss function to calculate the segmentation loss of the first branch network, using the IoULoss loss function to calculate the edge detection loss of the second branch network, performing weighted summation of the segmentation loss and the edge detection loss to obtain a joint loss value; Step S8: updating the CNN model according to the joint loss value; Step S9: If the number of iterations reaches the preset number, stop training, otherwise repeat steps S3 to S9.

[0009] Furthermore, the calculation process of the rainfall coefficient includes: Using a box counting method to calculate the fractal dimension of the edge of the rain cloud area edge map, to obtain the edge roughness coefficient; Calculating the variance of the grayscale values ​​of the pixels in the rain cloud area map to obtain the internal density coefficient; Calculating the vertical height difference between the cloud top and the cloud bottom using the rain cloud vertical profile data, calculating the thickness change rate for the vertical height difference at consecutive time intervals, and obtaining the rain cloud thickness coefficient; The edge roughness coefficient, the internal density coefficient and the rain cloud thickness coefficient are weightedly summed to obtain the rainfall coefficient.

[0010] Furthermore, the calculation process of the rain cloud influence coefficient includes: The optical flow method is used to calculate the motion vector field of each pixel point for the rain cloud area map at continuous time intervals, wherein the direction of the motion vector field is the moving direction of the rain cloud, and the average modulus length of the motion vector field is the moving speed of the rain cloud; The shortest distance is calculated using the Euclidean formula according to the coordinates of the rain cloud and the coordinates of the water conservancy project area; Calculating the cosine value of the angle between the rain cloud moving direction and the direction from the central rain cloud to the water conservancy project area, and multiplying the cosine value by the rain cloud moving speed to obtain a second rain cloud moving speed; The shortest distance is divided by the second rain cloud moving speed to calculate the rain cloud influence coefficient.

[0011] Furthermore, the calculation process of the vulnerability coefficient includes: Inputting the reservoir historical data into the basic model for training to obtain a trained deep learning model; wherein the reservoir historical data includes historical water level data, meteorological data, and reservoir inflow and outflow; Input the real-time data of the reservoir into the trained deep learning model for prediction to obtain the future water level of the reservoir; Calculate the ratio of the future water level of the reservoir to the water storage capacity of the reservoir to obtain the water storage coefficient of the reservoir; The number of flood occurrences and flood losses are obtained according to the historical data of flood defense, and the flood frequency and flood loss rate are obtained, and the flood frequency and the flood loss rate are multiplied to obtain the flood risk coefficient; Obtaining the slope and soil type according to the terrain and geographic data, and classifying the slope and soil type according to the terrain score to obtain a terrain score; calculating the average terrain score of the water conservancy project area according to the terrain score to obtain a terrain vulnerability coefficient; The weighted sum of building density and population density is used to obtain the social vulnerability coefficient; The reservoir storage coefficient, the flood risk coefficient, the terrain vulnerability coefficient and the social vulnerability coefficient are weightedly summed to obtain the vulnerability coefficient.

[0012] Furthermore, the calculation of the flood control coefficient includes: The rainfall coefficient, the rain cloud influence coefficient and the vulnerability coefficient are normalized respectively, and the normalized rainfall coefficient, the rain cloud influence coefficient and the vulnerability coefficient are weightedly summed to obtain the flood control coefficient.

[0013] Furthermore, the flood control model includes: Classifying the flood control coefficient into levels, including low risk level, medium risk level and high risk level; Minimizing the flood control coefficient and the flood control cost is used as a fitness function, and a genetic algorithm is used to perform the real-time regulation according to the fitness function and the scheduling constraints; wherein the real-time regulation includes conventional regulation and emergency regulation; If the flood control coefficient is the low risk level, the conventional regulation is performed; if the flood control coefficient is the medium risk level, the conventional regulation is performed and a warning is issued; if the flood control coefficient is the high risk level, the weight of the fitness function is adjusted and the emergency regulation is performed.

[0014] Furthermore, the scheduling constraints include: The maximum and minimum water level constraints for flood control are used to keep the reservoir water level within the maximum and minimum allowable range; Maximum and minimum release flow constraints are used to keep reservoir flow within the maximum and minimum allowable ranges; Outflow change constraint, used to keep the reservoir outflow change range within the allowable range; Water balance constraints, which are used to update reservoir storage based on outflow and inflow; The multi-reservoir coordination constraint is used to coordinate the outflow and inflow of the upstream and downstream reservoirs.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. By using the CNN model to perform image recognition of rain cloud areas, the spatial characteristics of rain clouds can be extracted, and the edge roughness coefficient, internal density coefficient and rain cloud thickness coefficient can be calculated, which improves the accuracy of rainfall prediction and makes weather monitoring more reliable. In addition, by using the optical flow method to analyze the movement speed and direction of rain clouds, the dynamics of rain clouds can be grasped in real time, and their potential impact on the water conservancy project system can be evaluated in time, providing a scientific basis for scheduling decisions, thereby effectively reducing flood risks and improving the flexibility of real-time scheduling of water conservancy project systems.

[0016] 2. By using deep learning models to predict future reservoir water levels, flood risks can be accurately assessed. Calculating flood risk coefficients in combination with historical flood defense data helps quantify the possible losses caused by floods. By analyzing the terrain vulnerability coefficient related to topography and the social vulnerability coefficient related to buildings and population density, the sensitivity of the system to floods can be fully reflected. Multi-dimensional assessments can help develop more targeted flood control strategies and improve the ability to respond to floods, thereby increasing the flexibility of real-time scheduling of water conservancy project systems.

[0017] 3. By calculating the flood control coefficient and establishing a flood control model based on the flood control coefficient, different risk levels are effectively distinguished, ensuring that strategies can be quickly adjusted and dispatched at different risks, thereby improving the flexibility and response speed to sudden flood events. In addition, combined with dispatch constraints such as water level, flow and water balance, the safety and effectiveness of the dispatch process are ensured. This real-time control method not only improves the scientificity and flexibility of water conservancy project system management, but also reduces flood risks and potential losses. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A schematic diagram of a method flow chart provided by an embodiment of the present invention; Figure 2 A schematic diagram of the CNN model training process provided by an embodiment of the present invention; Figure 3 A schematic diagram of the influence coefficient calculation structure provided by an embodiment of the present invention; Figure 4 A schematic diagram of the flood control coefficient calculation structure provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] See also Figures 1 to 4 The present invention provides a real-time dispatching method for a water conservancy project system based on a neural network technology, and the technical scheme is as follows: With climate change and the frequent occurrence of extreme weather events, the regulation of water conservancy project systems has become more complicated. Especially in the rainy season, there is often a flood risk caused by the superposition of local heavy rain and upstream water. The traditional scheduling method that relies on historical data and experience is difficult to cope with sudden hydrological changes and cannot achieve rapid scheduling of real-time rain cloud dynamics, water levels and watersheds, resulting in inefficient flood control and water resources management. Although the existing real-time scheduling method has improved the response capability, the lack of comprehensive analysis of rain cloud dynamics, terrain and social factors has resulted in limited flexibility in scheduling strategies.

[0021] In order to deal with these problems, a real-time dispatching method for water conservancy engineering systems based on neural network technology is proposed. This method can accurately predict rainfall, flood trends and future water level changes by acquiring satellite cloud images, rainfall data, reservoir water levels, river flow and other information in real time, combined with deep learning models. In addition, this method also formulates corresponding dispatching plans according to different flood control risk levels, realizes flexible dispatching under low, medium and high risk levels, ensures the reasonable allocation of regional water use, improves the efficiency of water resource utilization, avoids unnecessary flood discharge or waste of water resources, and thus improves the flexibility of system regulation.

[0022] Embodiment 1 Region A is facing water resource pressure caused by growing population and economic activities. It is also affected by frequent extreme rainfall events caused by climate change, which leads to increased flood risk. The original water conservancy engineering system cannot respond to sudden flood disasters and water resource allocation in a timely manner. For the water conservancy engineering system in region A, a real-time scheduling method for water conservancy engineering system based on neural network technology is proposed, aiming to improve the flexibility of real-time regulation of the system. Figure 1 As shown, including: Real-time data of a water conservancy project system is acquired at fixed time intervals, the real-time data including satellite cloud images, rain cloud vertical profile data and flood control data, and the satellite cloud images are preprocessed to obtain a first image data set.

[0023] Furthermore, the flood control data include: flood defense historical data, reservoir historical data, reservoir real-time data, topographic and geographical data, building density, population density and hydrological facilities.

[0024] Specifically, the fixed time interval can be set to 30 minutes, and meteorological satellites are used to transmit satellite cloud images in real time, and satellite cloud images are obtained in combination with ground receiving stations or special meteorological service platforms; ground radars, meteorological sondes and other equipment are used to obtain vertical profile data of rain clouds, including cloud thickness and density at different heights, so as to timely grasp the latest situation of rain clouds and improve the dynamic response capability of flood control and water resources scheduling; flood control data are obtained through historical databases, river basin monitoring stations and online sensing systems. Flood defense historical data include past flood events, maximum flood peak flow, flood impact range and losses, etc. Reservoir historical data include changes in reservoir water storage, scheduling records over the years and flood peak response measures, etc. Reservoir real-time data include real-time water level, inflow and outflow flow, current water storage and flood discharge status, etc. Topographic and geographical data are obtained through remote sensing technology or previous engineering records, and building density, population density and hydrological facilities are obtained through satellite images, urban planning data and water conservancy department databases. By collecting this data, we can comprehensively assess the impact of floods on regional infrastructure and residents, and comprehensively analyze the current conditions of rain clouds, river basins and reservoirs, providing an accurate basis for flood control scheduling, thereby improving the flexibility of system regulation.

[0025] Furthermore, the preprocessing includes: Performing color space conversion on the satellite cloud image to obtain a first cloud image; Specifically, the original format of satellite cloud images uses RGB or other color spaces. By converting the image into a color space suitable for cloud recognition and analysis, the characteristics of the cloud layer can be better highlighted. For example, converting an RGB satellite image containing clouds of different thicknesses into HSV space can enhance the brightness difference of the clouds, which is convenient for subsequent recognition algorithm processing.

[0026] Correcting the illumination and radiation of the first cloud image to obtain a second cloud image; Specifically, due to the lighting conditions and atmospheric radiation during satellite imaging, the clouds in satellite images may be affected by uneven lighting. For example, in the images taken by the satellite, the clouds in the east are brighter, while the clouds in the west are darker due to the angle of the sunlight. Through correction, the brightness of the entire image can be made uniform, reducing the errors caused by light differences.

[0027] The downsampling ratio is determined according to the task requirements, and bilinear interpolation is used to downsample the second cloud image according to the downsampling ratio to obtain a third cloud image, which can reduce the number of pixels while maintaining the clarity and details of the image as much as possible.

[0028] The third cloud image is subjected to noise removal using Gaussian filtering to obtain the first image data set, which can smooth the image so that the edge and structure of the cloud layer are clearer during subsequent image detection.

[0029] By preprocessing the images, not only do the satellite cloud images become higher quality, less noisy and easier to process, but the system processing speed and analysis accuracy are also improved, thereby improving the reliability and flexibility of flood control coefficient calculation and water conservancy scheduling.

[0030] Furthermore, the first image data set is subjected to image recognition using a CNN model, and the CNN model is used to extract spatial features of rain clouds, and obtain a rain cloud region map and a rain cloud region edge map. The image recognition process is as follows: feature extraction is performed based on the first image data set to obtain low-level image features; a first branch network and a second branch network are formed based on the low-level image features, the first branch network is used to segment the rain cloud region and output the rain cloud region map, and the second branch network is used to extract the edge of the rain cloud region and output the rain cloud region edge map; Furthermore, if Figure 2 As shown, the training steps of the CNN model include: Step S1: Acquire historical satellite cloud images and perform the preprocessing to obtain a historical cloud image data set; Step S2: label the image labels of the historical cloud image dataset to obtain a second image dataset, and input the second image dataset into the CNN model in the form of [H, W, C], where H is the height of the image, W is the width, and C is the number of channels; wherein the image labels include rain cloud areas, non-rain cloud areas, edge points, and non-edge points; Step S3: using a convolution layer to perform a convolution operation on the second image data set, using a convolution kernel of size 3×3 to slide on the cloud map, and outputting a first feature map; Step S4: using a ReLU activation function to set the negative value part of the convolution output to 0, using a pooling layer to perform a pooling operation on the first feature map, and outputting a second feature map after performing dimensionality reduction processing; Step S5: using the first branch network to perform a segmentation operation on the second feature map, each pixel is classified as "belonging to the rain cloud area" or "not belonging to the rain cloud area", and the rain cloud area map is output; Step S6: using the second branch network to perform edge feature detection on the second feature map, each pixel point is marked as an “edge point” or a “non-edge point”, and the rain cloud area edge map is output; Step S7: Use the cross entropy loss function to calculate the segmentation loss of the first branch network, use the IoULoss loss function to calculate the edge detection loss of the second branch network, and perform weighted summation of the segmentation loss and the edge detection loss to obtain a joint loss value, which is expressed as: ; in, is the segmentation loss value, is the cross entropy loss function, is the actual segmentation label, is the predicted output after model segmentation; is the edge loss value, is the IoU Loss loss function, is the actual label of the edge, is the predicted marginal output; is the final loss value, and are the hyperparameters of the two task weights, which can be set to 0.5 and 0.5 respectively.

[0031] Step S8: According to the joint loss value, the weight of the CNN model is updated using a back propagation algorithm to gradually improve the accuracy of rain cloud area recognition and edge detection; Step S9: If the number of iterations reaches the preset number, stop training, otherwise repeat steps S3 to S9.

[0032] Among them, the preset number of times can be set to 500 times.

[0033] Through the multi-branch network structure of the CNN model, after training 500 times, the final loss value was 0.2750, which was able to accurately identify the edge and area of ​​rain clouds. It not only improved the processing efficiency of meteorological data, but also provided strong data support for the real-time scheduling and optimization of water conservancy engineering systems, thereby improving the ability to respond to climate change and extreme weather events and the flexibility of system regulation.

[0034] like Figure 3 As shown, the edge roughness coefficient and the internal density coefficient are calculated based on the rain cloud area map and the rain cloud area edge map; the rain cloud thickness coefficient is calculated based on the rain cloud vertical profile data, and the rainfall coefficient is calculated based on the edge roughness coefficient, the internal density coefficient and the rain cloud thickness coefficient.

[0035] Furthermore, the calculation process of the rainfall coefficient includes: Using a box counting method to calculate the fractal dimension of the edge of the rain cloud area edge map, to obtain the edge roughness coefficient; Specifically, the edge map of the rain cloud area is converted into a binary image, with the edge point value being 1 and the rest being 0. Multiple square boxes of different sizes, such as 1×1, 2×2 pixels, etc., are used to cover the entire image using grids of different sizes. The number of boxes covered by at least one edge point is calculated, and the fractal dimension of the edge is calculated using the fractal dimension formula, which is expressed as: ; in, is the fractal dimension, is the side length of the box, is the number of covered boxes; the fractal dimension is used to describe the complexity of the shape. The larger the fractal dimension, the more complex the edge of the rain cloud and the higher the roughness.

[0036] The edge roughness coefficient is calculated from the fractal dimension of the edge and is expressed as: ; in, is the edge roughness coefficient, is the smallest fractal dimension of all rain cloud edges (such as a straight line), is the maximum fractal dimension of all rain cloud edges (like a fractal graphic).

[0037] The variance of the grayscale values ​​of the pixels in the rain cloud area map is calculated to obtain the internal density coefficient.

[0038] Specifically, if the rain cloud area map is a color image, it needs to be converted into a grayscale image and the variance of the grayscale value is calculated, which is expressed as: ; in, is the total number of pixels in the rain cloud area, is the gray value of the i-th pixel, is the mean of the grayscale values, is the variance of grayscale value; The internal density coefficient is obtained according to the variance of the gray value, which is expressed as: ; in, is the internal density coefficient, is the maximum variance of grayscale values; The vertical height difference between the cloud top and the cloud base is calculated using the rain cloud vertical profile data, and the thickness change rate is calculated for the vertical height difference at continuous time intervals to obtain the rain cloud thickness coefficient.

[0039] Specifically, the cloud top height and cloud bottom height are obtained using the rain cloud vertical profile data, and the vertical height difference is calculated and expressed as: ; in, is the vertical height difference, is the cloud top height, is the cloud base height; The thickness change rate is expressed as: ; in, is the thickness change rate, for The vertical height difference at that moment, for The vertical height difference at the moment; After normalizing the thickness change rate, the rain cloud thickness coefficient is obtained, which is expressed as: ; in, is the rain cloud thickness coefficient, is the maximum rate of change of height difference; The edge roughness coefficient, the internal density coefficient and the rain cloud thickness coefficient are weighted and summed to obtain the rainfall coefficient, which is expressed as: ; in, is the rainfall coefficient, , and is the weight parameter of each coefficient, which can be set to 0.3, 0.3 and 0.4. The higher the complexity of the edge, the stronger the meteorological activity. By calculating the edge roughness coefficient, the cloud edge morphology information can be provided as a potential indicator of rainfall intensity and duration. The internal density coefficient reflects the concentration and distribution characteristics of water vapor in the cloud, which further helps to analyze the possibility of rainfall. In addition, the rain cloud thickness change rate can evaluate the vertical thickness change of the rain cloud to capture the evolution of its morphology. The comprehensive evaluation of these three coefficients can more accurately describe the possibility of rainfall, thereby improving the flexibility of system regulation.

[0040] The optical flow method is used to calculate the rain cloud area map at continuous time intervals to obtain the rain cloud movement speed and rain cloud movement direction, and the shortest distance between the central rain cloud in the rain cloud area map and the water conservancy project area is calculated; the rain cloud influence coefficient is calculated based on the rain cloud movement speed, the rain cloud movement direction and the shortest distance.

[0041] Furthermore, the calculation process of the rain cloud influence coefficient includes: The optical flow method is used to calculate the motion vector field of each pixel point for the rain cloud area map at continuous time intervals, wherein the direction of the motion vector field is the moving direction of the rain cloud, and the average modulus length of the motion vector field is the moving speed of the rain cloud; According to the coordinates of the rain cloud and the coordinates of the water conservancy project area, the shortest distance is calculated using the Euclidean formula and is expressed as: ; in, is the shortest distance, is the center coordinate of the rain cloud, are the center coordinates of the water conservancy project area; The cosine value of the angle between the moving direction of the rain cloud and the direction from the central rain cloud to the water conservancy project area is calculated, and the cosine value is multiplied by the moving speed of the rain cloud to obtain the second moving speed of the rain cloud; the shortest distance is divided by the second moving speed of the rain cloud to calculate the rain cloud influence coefficient, which is expressed as: ; in, is the rain cloud influence coefficient, is the moving speed of rain clouds, is the cosine value of the angle between the moving direction of the rain cloud and the direction of the water conservancy project area, It is the angle between the moving direction of rain clouds and the direction of the water conservancy project area. By calculating the rain cloud impact coefficient, the potential impact of rain clouds on the water conservancy project area can be dynamically reflected, especially when dealing with sudden rainfall, which can improve the flexibility of system regulation, thereby improving the system's early warning capability and dispatch response efficiency.

[0042] The future water level of the reservoir is predicted based on the flood control data using a deep learning model, and the vulnerability coefficient is calculated based on the flood control data and the future water level of the reservoir.

[0043] Furthermore, the calculation process of the vulnerability coefficient includes: Inputting the reservoir historical data into the basic model for training to obtain a trained deep learning model; wherein the reservoir historical data includes historical water level data, meteorological data, and reservoir inflow and outflow; Input the real-time data of the reservoir into the deep learning model for prediction to obtain the future water level of the reservoir; Among them, the deep learning model can choose LSTM model, GRU model or CNN model; Specifically, in the training phase, the actual water level is used as the label and the mean square error is used as the loss function to measure the deviation between the actual water level and the predicted water level. The model continuously adjusts the parameters until the error between its prediction result and the actual water level data is minimized. After 1,000 trainings, the error remains at 0.17. In the prediction phase, the model can output the water level changes in the future through calculations, for example, the water level will reach 150.5 meters tomorrow.

[0044] Calculate the ratio of the future water level of the reservoir to the water storage capacity of the reservoir to obtain the water storage coefficient of the reservoir; Specifically, for example, the maximum water storage capacity of the reservoir is 50 million cubic meters, and the water storage capacity corresponding to the predicted water level is 30 million cubic meters, then the reservoir water storage coefficient is 0.6.

[0045] The number of flood occurrences and flood losses are obtained according to the historical data of flood defense, and the flood frequency and flood loss rate are obtained, and the flood frequency and the flood loss rate are multiplied to obtain the flood risk coefficient; Specifically, for example, in the history of a certain region, a water conservancy project area has experienced five floods in 20 years, with a total loss of 50 million yuan and a total economic value of 20 billion yuan. After calculation, the flood risk coefficient can be obtained as 0.000625. If the calculated value is too small, the value can be enlarged according to the theoretical upper limit of the risk coefficient of the region, that is, the product of the maximum total flood loss and the maximum flood loss rate. For example, if the theoretical upper limit is 0.01, the flood risk coefficient is 0.0625.

[0046] Obtaining the slope and soil type according to the terrain and geographic data, and classifying the slope and soil type according to the terrain score to obtain a terrain score; calculating the average terrain score of the water conservancy project area according to the terrain score to obtain a terrain vulnerability coefficient; Specifically, the division method can be: If the slope is less than 5% or the soil type is low vulnerability, a score of 0.3 is assigned; if the slope is less than 15% and the slope is greater than 5%, or the soil type is medium vulnerability, a score of 0.6 is assigned; if the slope is greater than 15% or the soil type is high vulnerability, a score of 1 is assigned.

[0047] The weighted sum of building density and population density is used to obtain the social vulnerability coefficient; Among them, the building density and population density are normalized separately, and the weights can be set to 0.4 and 0.6 respectively.

[0048] The reservoir storage coefficient, the flood risk coefficient, the terrain vulnerability coefficient and the social vulnerability coefficient are weighted and summed to obtain the vulnerability coefficient, which is expressed as: ; in, is the vulnerability coefficient, is the reservoir storage coefficient, is the flood risk factor, is the terrain vulnerability coefficient, is the social vulnerability coefficient, , , and are the weights of each coefficient, which can all be set to 0.2.

[0049] By predicting the future water level of the reservoir through a deep learning model and calculating the vulnerability coefficient by combining flood control data, terrain data and social data, it is possible to provide potential risk information of the reservoir, which will help the water conservancy department take measures in advance, improve flood control and disaster resistance capabilities, and ensure the safety of the reservoir and surrounding areas, thereby improving the scientific nature of reservoir management and the accuracy of real-time scheduling.

[0050] A flood control coefficient is calculated based on the rainfall coefficient, the rain cloud influence coefficient and the vulnerability coefficient, and a flood control model is established based on the flood control coefficient. The flood control model is used to perform real-time control on the water conservancy project system.

[0051] Furthermore, if Figure 4 As shown, the calculation of the flood control coefficient includes: The rainfall coefficient, the rain cloud influence coefficient and the vulnerability coefficient are normalized respectively, and the normalized rainfall coefficient, the rain cloud influence coefficient and the vulnerability coefficient are weighted summed to obtain the flood control coefficient, which is expressed as: ; in, is the flood control factor, is the rainfall coefficient, is the rain cloud influence coefficient, is the vulnerability coefficient, , and are the weights of each coefficient, which can be set to 0.4, 0.3 and 0.3 respectively.

[0052] By combining rainfall intensity, rain cloud impact and the vulnerability of the region, the calculation of the flood control coefficient is more accurate, and the possible flood risk can be predicted more accurately, thereby improving the flexibility of system regulation. In addition, by setting weights, it can be adjusted according to different regional characteristics, climate patterns and infrastructure. For example, areas with more rainfall can increase the weight of the rainfall coefficient, while areas with higher social vulnerability can pay more attention to the vulnerability coefficient, thereby more adapting to flood control plans in different regions.

[0053] Furthermore, the flood control model includes: Classifying the flood control coefficient into levels, including low risk level, medium risk level and high risk level; Specifically, the risk level can be set according to the value range of the flood control coefficient. Here, the value range of the flood control coefficient is 0-1, and the risk level is: Low risk level ( ): The reservoir is operating normally and the risk is low.

[0054] Medium risk level ( ): The reservoir has a certain flood risk and the discharge volume needs to be increased appropriately.

[0055] High risk level ( ): The reservoir faces a high flood risk and emergency measures must be taken immediately, such as rapid water release and increased dispatch frequency.

[0056] Minimizing the flood control coefficient and the flood control cost is used as a fitness function, and a genetic algorithm is used to perform the real-time regulation according to the fitness function and the scheduling constraints; wherein the real-time regulation includes conventional regulation and emergency regulation; If the flood control coefficient is the low risk level, the conventional regulation is performed; if the flood control coefficient is the medium risk level, the conventional regulation is performed and a warning is issued; if the flood control coefficient is the high risk level, the weight of the fitness function is adjusted and the emergency regulation is performed.

[0057] Specifically, the fitness function can be expressed as: ; in, For fitness, is the normalized flood control cost, and are the weight coefficients of the two objectives respectively.

[0058] Among them, when the risk level is high, increasing the priority of emergency regulation and reducing the weight of cost can Set to 0.95, Set to 0.05; at low and medium risk levels, Set to 0.7, Set to 0.3.

[0059] In order to ensure that the scheduling plan meets all constraints, the violation of the constraints can be added to the fitness function, and the expanded fitness function is: ; in, There is a penalty function corresponding to each constraint. The greater the violation, the greater the penalty.

[0060] Specifically, the genetic algorithm implementation steps are: First, an initial population is generated with a population size of 30. Each chromosome represents a reservoir operation scheme. The chromosome includes operation parameters, such as the discharge and inflow at each time step. Calculate the fitness function of each chromosome. The higher the fitness value, the more the scheduling plan of the chromosome meets the requirements. Use roulette wheel to select chromosomes to generate new next generation chromosomes. The chromosomes with higher fitness have a greater chance of being selected. Exchange the scheduling parameters of two chromosomes in different time periods to generate new chromosomes. The probability can be set to 0.8. Randomly adjust the scheduling parameters of certain time periods to increase the diversity of the population. The probability can be set to 0.05; Generate new populations through operations such as selection, crossover, and mutation; If the number of iterations reaches 800, the iteration is terminated and the scheduling parameters are output.

[0061] By taking the minimization of flood control coefficient and flood control cost as the fitness function, both flood control safety and economic cost are taken into account. In addition, by combining conventional control and emergency control schemes, real-time adjustments can be made according to the flood control coefficient, which improves the flexibility and safety of scheduling and avoids the risk of extreme weather or sudden floods. Finally, the effective combination of hierarchical scheduling and genetic algorithms can achieve efficient and safe regulation of the water conservancy system and improve the flexibility of system regulation.

[0062] Furthermore, the scheduling constraints include: The maximum and minimum water level constraints for flood control are used to keep the reservoir water level within the maximum and minimum allowable ranges, expressed as: ; in, is the water level at time t, is the minimum permissible water level, which can be set to 40m. It is the maximum permissible water level and can be set to 80m.

[0063] The maximum and minimum flow constraints of discharge are used to keep the reservoir flow within the maximum and minimum allowed ranges, expressed as: ; in, is the outbound flow at time t, is the minimum outbound flow rate, which can be set to 100 cubic meters per second. It is the maximum outbound flow rate, which can be set to 400 cubic meters per second.

[0064] The outflow change constraint is used to keep the outflow change range of the reservoir within the allowable range, which is expressed as: ; in, is the flow change between adjacent time steps, is the outbound flow at time t-1, is the maximum permissible flow rate change and can be set to 50 m3 / s.

[0065] The water balance constraint is used to update the reservoir storage according to the outflow and inflow of the reservoir, which is expressed as: ; in, is the storage capacity at time t+1, is the storage capacity at time t, is the inflow flow at time t.

[0066] The multi-reservoir coordination constraint is used to coordinate the outflow and inflow of the upstream and downstream reservoirs, and is expressed as: ; in, is the discharge of the upstream reservoir at time t, is the inflow of the current reservoir at time t, It is the inflow carrying capacity of the downstream reservoir and can be set to 300 cubic meters per second.

[0067] Introducing these constraints in real-time scheduling not only strengthens the cooperation between upstream and downstream reservoirs and improves regional flood control capabilities, but also ensures that the water level and flow of the reservoir remain within a safe range, effectively suppresses the drastic fluctuations in outflow flow, avoids the risks of overflow and drying up, and thus improves the flexibility of regulation.

[0068] Embodiment 2 Region B has abundant water resources, but due to its complex terrain and diverse climate, water resource management faces challenges. With the impact of climate change, Region B has also experienced extreme weather events such as drought and heavy rain, resulting in uneven distribution of water resources and ecosystem degradation. Therefore, the water conservancy engineering system in Region B needs a new regulation method to improve the flexibility of regulation. A real-time scheduling method for water conservancy engineering system based on neural network technology, including: Acquire real-time data of a water conservancy project system at fixed time intervals, the real-time data including satellite cloud images, rain cloud vertical profile data and flood control data, and pre-process the satellite cloud images to obtain a first image data set; Among them, the time interval can be set to 1h; The first image data set is subjected to image recognition using a CNN model, wherein the CNN model is used to extract spatial features of rain clouds, and obtain a rain cloud region map and a rain cloud region edge map; the image recognition process is as follows: feature extraction is performed according to the first image data set to obtain low-level image features; a first branch network and a second branch network are formed according to the low-level image features, wherein the first branch network is used to segment the rain cloud region and output the rain cloud region map, and the second branch network is used to extract the edge of the rain cloud region and output the rain cloud region edge map; An edge roughness coefficient and an internal density coefficient are calculated based on the rain cloud area map and the rain cloud area edge map; a rain cloud thickness coefficient is calculated based on the rain cloud vertical profile data, and a rainfall coefficient is calculated based on the edge roughness coefficient, the internal density coefficient and the rain cloud thickness coefficient; The rain cloud area map at continuous time intervals is calculated using the optical flow method to obtain the rain cloud moving speed and the rain cloud moving direction, and the shortest distance between the central rain cloud in the rain cloud area map and the water conservancy project area is calculated; the rain cloud influence coefficient is calculated according to the rain cloud moving speed, the rain cloud moving direction and the shortest distance; Predicting the future water level of the reservoir using a deep learning model based on the flood control data, and calculating based on the flood control data and the future water level of the reservoir to obtain a vulnerability coefficient; A flood control coefficient is calculated based on the rainfall coefficient, the rain cloud influence coefficient and the vulnerability coefficient, and a flood control model is established based on the flood control coefficient. The flood control model is used to perform real-time control on the water conservancy project system.

[0069] Furthermore, the flood control data include: flood defense historical data, reservoir historical data, reservoir real-time data, topographic and geographical data, building density, population density and hydrological facilities.

[0070] Furthermore, the preprocessing includes: Obtain historical satellite cloud image data of region B in the past five years, perform color space conversion on the satellite cloud image, convert the original cloud image from RGB color space to HSV color space, and obtain a first cloud image; Correcting the illumination and radiation of the first cloud image to ensure the illumination consistency of different areas in the image, thereby obtaining a second cloud image; Determine the downsampling ratio according to the task requirements. Here, the downsampling ratio is set to 0.5, and use bilinear interpolation to downsample the second cloud image according to the downsampling ratio to reduce the calculation burden, thereby obtaining a third cloud image. The third cloud image is subjected to noise removal using Gaussian filtering, and the filter size is set to 5×5 to improve image quality, thereby obtaining the first image data set.

[0071] Furthermore, the training steps of the CNN model include: Step S1: Acquire historical satellite cloud images and perform the preprocessing to obtain a historical cloud image data set; Step S2: labeling the historical cloud image dataset with image labels to obtain a second image dataset, and inputting the second image dataset into the CNN model; wherein the image labels include rain cloud areas, non-rain cloud areas, edge points, and non-edge points, so as to facilitate subsequent model training; Step S3: Use a convolution layer to perform a convolution operation on the second image data set, set the convolution kernel size to 3×3, the stride to 1, and output a first feature map; Step S4: Use the ReLU activation function and the pooling layer to perform a pooling operation on the first feature map, with a pooling window of 2×2 and a stride of 2, and output a second feature map; Step S5: using the first branch network to perform a segmentation operation on the second feature map, and outputting the rain cloud area map; Step S6: using the second branch network to perform edge feature detection on the second feature map, and outputting the rain cloud area edge map; Among them, the first branch network and the second branch network are fully convolutional networks; Step S7: using the cross entropy loss function to calculate the segmentation loss of the first branch network, using the IoULoss loss function to calculate the edge detection loss of the second branch network, performing weighted summation of the segmentation loss and the edge detection loss to obtain a joint loss value; Among them, the segmentation loss weight is set to 0.7 and the edge detection loss weight is set to 0.3.

[0072] Step S8: updating the CNN model according to the joint loss value; Step S9: If the number of iterations reaches the preset number, the maximum number of iterations of training is set to 1000, and the training is stopped, otherwise the steps S3 to S9 are repeated.

[0073] Furthermore, the calculation process of the rainfall coefficient includes: Using a box counting method to calculate the fractal dimension of the edge of the rain cloud area edge map, to obtain the edge roughness coefficient; Among them, the edge roughness coefficient can be directly equal to the fractal dimension of the edge; Calculating the variance of the grayscale values ​​of the pixels in the rain cloud area map to obtain the internal density coefficient; Specifically, assuming that the grayscale value of the rain cloud area map is [120, 130, 140, 135, 125, 128, 132, 138, 142, 134], the average grayscale value is calculated to be 131.5, and the grayscale value variance is 50.25.

[0074] Calculating the vertical height difference between the cloud top and the cloud bottom using the rain cloud vertical profile data, calculating the thickness change rate for the vertical height difference at consecutive time intervals, and obtaining the rain cloud thickness coefficient; Specifically, assuming that the cloud top height is 3000m and the cloud bottom height is 1500m, the vertical height difference is 1500m. If the height difference before and after within 1 hour is 1500m and 1200m respectively, the change rate is 300m / h.

[0075] The edge roughness coefficient, the internal density coefficient and the rain cloud thickness coefficient are normalized and then weighted summed to obtain the rainfall coefficient.

[0076] Among them, the weights are set to 0.4, 0.4 and 0.2 respectively.

[0077] Furthermore, the calculation process of the rain cloud influence coefficient includes: The optical flow method is used to calculate the motion vector field of each pixel point for the rain cloud area map at continuous time intervals, wherein the direction of the motion vector field is the moving direction of the rain cloud, and the average modulus length of the motion vector field is the moving speed of the rain cloud; Specifically, the motion vector of each pixel is expressed as (dx, dy), which are the changes in the horizontal and vertical directions respectively. The calculation formula for the average modulus is: ; in, is the average modulus length, is the number of pixels, i is the pixel sequence number; The shortest distance is calculated using the Euclidean formula according to the coordinates of the rain cloud and the coordinates of the water conservancy project area; Calculating the cosine value of the angle between the rain cloud moving direction and the direction from the central rain cloud to the water conservancy project area, and multiplying the cosine value by the rain cloud moving speed to obtain a second rain cloud moving speed; The shortest distance is divided by the second rain cloud moving speed to calculate the rain cloud influence coefficient.

[0078] Furthermore, the calculation process of the vulnerability coefficient includes: The reservoir historical data of the past five years was input into the basic model for training to obtain a trained deep learning model. The prediction accuracy of the model on the test set reached 87%; the reservoir historical data included historical water level data, meteorological data, and reservoir inflow and outflow; Input the real-time data of the reservoir into the trained deep learning model for prediction to obtain the future water level of the reservoir; Calculate the ratio of the future water level of the reservoir to the water storage capacity of the reservoir to obtain the water storage coefficient of the reservoir; The number of flood occurrences and flood losses are obtained according to the historical data of flood defense, and the flood frequency and flood loss rate are obtained, and the flood frequency and the flood loss rate are multiplied to obtain the flood risk coefficient; Obtaining the slope and soil type according to the terrain and geographic data, and classifying the slope and soil type according to the terrain score to obtain a terrain score; calculating the average terrain score of the water conservancy project area according to the terrain score to obtain a terrain vulnerability coefficient; The weighted sum of building density and population density is used to obtain the social vulnerability coefficient; The reservoir storage coefficient, the flood risk coefficient, the terrain vulnerability coefficient and the social vulnerability coefficient are weightedly summed to obtain the vulnerability coefficient.

[0079] Among them, the weights are set to 0.25, 0.35, 0.2, and 0.2.

[0080] Furthermore, the calculation of the flood control coefficient includes: The rainfall coefficient, the rain cloud influence coefficient and the vulnerability coefficient are normalized respectively, and the normalized rainfall coefficient, the rain cloud influence coefficient and the vulnerability coefficient are weightedly summed to obtain the flood control coefficient, which provides a quantitative basis for flood control management in area B and facilitates the formulation of effective flood control measures and emergency plans.

[0081] Among them, the weights are set to 0.4, 0.3, and 0.3.

[0082] Furthermore, the flood control model includes: Classifying the flood control coefficient into levels, including low risk level, medium risk level and high risk level; Among them, the risk level is: low risk level ( ), medium risk level ( ), high risk level ( ), assuming the calculated flood control coefficient is 0.6982, it is within the medium risk level.

[0083] Taking minimization of the flood control coefficient and the flood control cost as a fitness function, and using a genetic algorithm to perform the real-time regulation according to the fitness function and the scheduling constraints; The real-time water conservancy regulation includes conventional regulation and emergency regulation. The parameters of the genetic algorithm are set as follows: population size is 100, selection probability is 40%, crossover probability is 80%, mutation probability is 5%, and termination condition is the maximum number of iterations is 1000 generations.

[0084] If the flood control coefficient is the low risk level, the conventional regulation is performed; if the flood control coefficient is the medium risk level, the conventional regulation is performed and a warning is issued; if the flood control coefficient is the high risk level, the weight of the fitness function is adjusted and the emergency regulation is performed.

[0085] Among them, at high risk levels, the flood control coefficient is set to 0.9 and the cost weight is set to 0.1; at low and medium risk levels, the flood control coefficient is set to 0.5 and the cost weight is set to 0.5.

[0086] Furthermore, the scheduling constraints include: The maximum and minimum water level constraints for flood control are used to keep the reservoir water level within the maximum and minimum allowable range; Among them, the maximum water level of the reservoir is set at 100 meters and the minimum water level is set at 50 meters.

[0087] Maximum and minimum release flow constraints are used to keep reservoir flow within the maximum and minimum allowable ranges; Among them, the maximum discharge flow is 500 cubic meters / second and the minimum flow is 100 cubic meters / second.

[0088] Outflow change constraint, used to keep the reservoir outflow change range within the allowable range; Among them, the change range of outbound flow should be kept within 50 cubic meters per second.

[0089] Water balance constraints, which are used to update reservoir storage based on outflow and inflow; Among them, the water storage capacity is updated according to the reservoir's inflow (such as 600 cubic meters / second) and outflow (such as the current outflow is 300 cubic meters / second).

[0090] The multi-reservoir coordination constraint is used to coordinate the outflow and inflow of the upstream and downstream reservoirs.

[0091] Among them, the inflow carrying capacity of the downstream reservoir is set at 250 cubic meters per second.

[0092] Specifically, assuming that the flood control coefficient is at a medium risk level, the system performs routine regulation and issues a warning to inform relevant personnel to pay attention to possible flood risks, and then uses genetic algorithms to calculate the dispatch parameters, for example, the discharge is adjusted to 400 cubic meters per second. By combining genetic algorithms to optimize the real-time dispatch strategy of the water conservancy system, when the flood risk increases, the dispatch constraints can be used to provide effective decision-making support for water conservancy management, thereby improving the flexibility of system regulation.

[0093] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A real-time dispatching method for a water conservancy project system based on neural network technology, comprising: Features: Acquire real-time data of a water conservancy project system at fixed time intervals, the real-time data including satellite cloud images, rain cloud vertical profile data and flood control data, and pre-process the satellite cloud images to obtain a first image data set; Performing image recognition on the first image data set using a CNN model, wherein the CNN model is used to extract spatial features of rain clouds to obtain a rain cloud area map and a rain cloud area edge map; An edge roughness coefficient and an internal density coefficient are calculated based on the rain cloud area map and the rain cloud area edge map; a rain cloud thickness coefficient is calculated based on the rain cloud vertical profile data; a rainfall coefficient is calculated based on the edge roughness coefficient, the internal density coefficient and the rain cloud thickness coefficient; The rain cloud area map at continuous time intervals is calculated using the optical flow method to obtain the rain cloud moving speed and the rain cloud moving direction, and the shortest distance between the central rain cloud in the rain cloud area map and the water conservancy project area is calculated; the rain cloud influence coefficient is calculated according to the rain cloud moving speed, the rain cloud moving direction and the shortest distance; Predicting the future water level of the reservoir using a deep learning model based on the flood control data, and calculating a vulnerability coefficient based on the flood control data and the future water level of the reservoir; A flood control coefficient is calculated based on the rainfall coefficient, the rain cloud influence coefficient and the vulnerability coefficient, and a flood control model is established based on the flood control coefficient. The flood control model is used to perform real-time control on the water conservancy project system.

2. According to claim 1, a real-time scheduling method for water conservancy engineering system based on neural network technology, It is characterized in that The flood control data include: historical flood defense data, historical reservoir data, real-time reservoir data, topographic and geographical data, building density, population density and hydrological facilities.

3. According to claim 1, a real-time scheduling method for water conservancy engineering system based on neural network technology, It is characterized in that The pre-processing comprises: Performing color space conversion on the satellite cloud image to obtain a first cloud image; Correcting the illumination and radiation of the first cloud image to obtain a second cloud image; Determine a downsampling ratio according to task requirements, and use bilinear interpolation to downsample the second cloud image according to the downsampling ratio to obtain a third cloud image; The third cloud image is subjected to a Gaussian filter to remove noise, thereby obtaining the first image data set.

4. According to claim 1, a real-time scheduling method for water conservancy engineering system based on neural network technology, It is characterized in that The training steps of the CNN model include: Step S1: Acquire historical satellite cloud images and perform the preprocessing to obtain a historical cloud image data set; Step S2: labeling the image labels of the historical cloud image dataset to obtain a second image dataset, and inputting the second image dataset into the CNN model; wherein the CNN model includes a first branch network and a second branch network; the image labels include rain cloud areas, non-rain cloud areas, edge points, and non-edge points; Step S3: using a convolution layer to perform a convolution operation on the second image data set, and output a first feature map; Step S4: using a ReLU activation function and a pooling layer to perform a pooling operation on the first feature map, and output a second feature map; Step S5: using the first branch network to perform a segmentation operation on the second feature map, and outputting the rain cloud area map; Step S6: using the second branch network to perform edge feature detection on the second feature map, and outputting the rain cloud area edge map; Step S7: using the cross entropy loss function to calculate the segmentation loss of the first branch network, using the IoU Loss loss function to calculate the edge detection loss of the second branch network, performing weighted summation of the segmentation loss and the edge detection loss to obtain a joint loss value; Step S8: updating the CNN model according to the joint loss value; Step S9: If the number of iterations reaches the preset number, stop training, otherwise repeat steps S3 to S9.

5. According to claim 1, a real-time scheduling method for water conservancy engineering system based on neural network technology, It is characterized in that The calculation process of the rainfall coefficient includes: Using a box counting method to calculate the fractal dimension of the edge of the rain cloud area edge map, to obtain the edge roughness coefficient; Calculating the variance of the grayscale values ​​of the pixels in the rain cloud area map to obtain the internal density coefficient; Calculating the vertical height difference between the cloud top and the cloud bottom using the rain cloud vertical profile data, calculating the thickness change rate for the vertical height difference at consecutive time intervals, and obtaining the rain cloud thickness coefficient; The edge roughness coefficient, the internal density coefficient and the rain cloud thickness coefficient are weightedly summed to obtain the rainfall coefficient.

6. A real-time dispatching method for a water conservancy project system based on neural network technology according to claim 1, It is characterized in that The calculation process of the rain cloud influence coefficient includes: The optical flow method is used to calculate the motion vector field of each pixel point for the rain cloud area map at continuous time intervals, wherein the direction of the motion vector field is the moving direction of the rain cloud, and the average modulus length of the motion vector field is the moving speed of the rain cloud; The shortest distance is calculated using the Euclidean formula according to the coordinates of the rain cloud and the coordinates of the water conservancy project area; Calculating the cosine value of the angle between the rain cloud moving direction and the direction from the central rain cloud to the water conservancy project area, and multiplying the cosine value by the rain cloud moving speed to obtain a second rain cloud moving speed; The shortest distance is divided by the second rain cloud moving speed to calculate the rain cloud influence coefficient.

7. A real-time dispatching method for a water conservancy project system based on neural network technology according to claim 1, It is characterized in that The calculation process of the vulnerability coefficient includes: Inputting the reservoir historical data into the basic model for training to obtain a trained deep learning model; wherein the reservoir historical data includes historical water level data, meteorological data, and reservoir inflow and outflow; Input the real-time data of the reservoir into the trained deep learning model for prediction to obtain the future water level of the reservoir; Calculate the ratio of the future water level of the reservoir to the water storage capacity of the reservoir to obtain the water storage coefficient of the reservoir; The number of flood occurrences and flood losses are obtained according to the historical data of flood defense, and the flood frequency and flood loss rate are obtained, and the flood frequency and the flood loss rate are multiplied to obtain the flood risk coefficient; Obtaining the slope and soil type according to the terrain and geographic data, and classifying the slope and soil type according to the terrain score to obtain a terrain score; calculating the average terrain score of the water conservancy project area according to the terrain score to obtain a terrain vulnerability coefficient; The weighted sum of building density and population density is used to obtain the social vulnerability coefficient; The reservoir water storage coefficient, the flood risk coefficient, the terrain vulnerability coefficient and the social vulnerability coefficient are weightedly summed to obtain the vulnerability coefficient.

8. According to claim 1, a real-time dispatching method for water conservancy engineering system based on neural network technology, It is characterized in that The calculation of the flood control coefficient includes: The rainfall coefficient, the rain cloud influence coefficient and the vulnerability coefficient are normalized respectively, and the normalized rainfall coefficient, the rain cloud influence coefficient and the vulnerability coefficient are weightedly summed to obtain the flood control coefficient.

9. The real-time dispatching method of a water conservancy project system based on neural network technology according to claim 1, It is characterized in that The flood control model includes: Classifying the flood control coefficient into levels, including low risk level, medium risk level and high risk level; Minimizing the flood control coefficient and the flood control cost is used as a fitness function, and a genetic algorithm is used to perform the real-time regulation according to the fitness function and the scheduling constraints; wherein the real-time regulation includes conventional regulation and emergency regulation; If the flood control coefficient is the low risk level, the conventional regulation is performed; if the flood control coefficient is the medium risk level, the conventional regulation is performed and a warning is issued; if the flood control coefficient is the high risk level, the weight of the fitness function is adjusted and the emergency regulation is performed.

10. A real-time dispatching method for a water conservancy project system based on neural network technology according to claim 9, It is characterized in that The scheduling constraints include: The maximum and minimum water level constraints for flood control are used to keep the reservoir water level within the maximum and minimum allowable range; Maximum and minimum release flow constraints are used to keep reservoir flow within the maximum and minimum allowable ranges; Outflow change constraint, used to keep the reservoir outflow change range within the allowable range; A water balance constraint, used to update the reservoir water storage capacity according to the outflow and inflow; The multi-reservoir coordination constraint is used to coordinate the outflow and inflow of the upstream and downstream reservoirs.

Images