A flood prevention and early warning method for reservoir irrigation areas based on big data analysis

Abstract

The present application provides a reservoir irrigation area flood prevention and warning method based on big data analysis, including: obtaining the monitoring data of the reservoir irrigation area to be predicted by flood forecasting, loading the monitoring data of the reservoir irrigation area into the pre-debugging characterization coding network component in the Hongxun prediction neural network for characterization coding, and obtaining the coded representation of the monitoring data; the pre-debugging characterization coding network component is obtained by pre-debugging the characterization coding network component in the Hongxun prediction neural network without labels based on the pre-debugging reservoir irrigation area monitoring database; performing Hongxun prediction on the coded representation of the monitoring data based on the Hongxun reasoning component in the Hongxun prediction neural network, and obtaining the Hongxun trigger score corresponding to the monitoring data of the reservoir irrigation area. The present application can improve the debugging quality of the Hongxun prediction neural network in this way, so that the prediction accuracy and reliability of the Hongxun prediction neural network in performing flood forecasting on the monitoring data of the reservoir irrigation area are higher.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a reservoir irrigation area flood prevention and early warning method based on big data analysis. Background Art

[0002] In the management of reservoir irrigation areas, the accuracy and reliability of flood warning systems are crucial to protecting people's lives and property. Traditional flood warning methods often rely on empirical judgment and mathematical models. These methods have limitations when processing large-scale, high-dimensional monitoring data, and it is difficult to accurately capture the complex hydrological dynamics and potential flood risks of reservoir irrigation areas. With the development of big data technology and artificial intelligence, flood warning systems based on machine learning have gradually become a research hotspot. Flood warning systems based on machine learning include steps such as data collection, feature extraction, model training, and prediction. However, these systems still face many challenges in practical applications. The performance of flood warning systems is highly dependent on the quality and quantity of training data. In the absence of sufficient labeled data, the generalization ability and prediction accuracy of the model are often limited. Parameter tuning during model training is crucial to improving prediction accuracy, but traditional parameter tuning methods often rely on manual experience and trial and error, which are inefficient and difficult to guarantee the optimal solution. Summary of the invention

[0003] In view of this, the present application provides a reservoir irrigation area flood prevention and early warning method based on big data analysis.

[0004] The technical solution of the present application is implemented as follows: a reservoir irrigation area flood prevention and warning method based on big data analysis, the method comprising: obtaining reservoir irrigation area monitoring data to be predicted for flood warning; loading the reservoir irrigation area monitoring data into the pre-debugging representation coding network component in the flood warning prediction neural network for representation coding to obtain a coding representation of the monitoring data; the pre-debugging representation coding network component is obtained by pre-debugging the representation coding network component in the flood warning prediction neural network without labels based on the pre-debugging reservoir irrigation area monitoring database; performing flood warning prediction on the coding representation of the monitoring data based on the flood warning reasoning component in the flood warning prediction neural network to obtain a flood warning trigger score corresponding to the reservoir irrigation area monitoring data. Among them, the adjustment process of the Hongxun prediction neural network includes: obtaining a pre-debugging reservoir irrigation area monitoring database, a reservoir irrigation area monitoring data template and a prior Hongxun trigger score corresponding to the reservoir irrigation area monitoring data template; performing unlabeled pre-debugging on the representation coding network component based on the reservoir irrigation area monitoring data in the pre-debugging reservoir irrigation area monitoring database to obtain the pre-debugging representation coding network component; performing Hongxun prediction on the reservoir irrigation area monitoring data template through the pre-debugging representation coding network component and the Hongxun reasoning component to obtain a predicted Hongxun trigger score; determining a prediction error based on the predicted Hongxun trigger score and the prior Hongxun trigger score; and adjusting the component parameters of the pre-debugging representation coding network component and / or the Hongxun reasoning component based on the prediction error.

[0005] Beneficial effects of the present application: The present application is based on obtaining a pre-debugging reservoir irrigation area monitoring database, a reservoir irrigation area monitoring data template and a prior flood signal trigger score corresponding to the reservoir irrigation area monitoring data template, and performs label-free pre-debugging of the representation coding network component based on the reservoir irrigation area monitoring data in the pre-debugging reservoir irrigation area monitoring database to obtain the pre-debugging representation coding network component, and performs flood signal prediction on the reservoir irrigation area monitoring data template through the pre-debugging representation coding network component and the flood signal reasoning component to obtain the predicted flood signal trigger score, and determines the prediction error based on the predicted flood signal trigger score and the prior flood signal trigger score, and adjusts the component parameters of the pre-debugging representation coding network component and / or the flood signal reasoning component according to the prediction error. In the debugging phase of the Hongxun prediction neural network, the representation coding network components in the Hongxun prediction neural network are pre-debugged without labels based on the pre-debugging reservoir irrigation area monitoring database with a large sample size. After obtaining the pre-debugging representation coding network components, the Hongxun prediction neural network is debugged again based on the reservoir irrigation area monitoring data template. This can improve the debugging quality of the Hongxun prediction neural network and make the Hongxun prediction neural network more accurate and reliable in making Hongxun predictions on the reservoir irrigation area monitoring data. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1A schematic diagram of the implementation process of a reservoir irrigation area flood prevention and early warning method based on big data analysis provided in an embodiment of the present application.

[0007] Figure 2 A schematic diagram of the calibration process of a Hongxun prediction neural network provided in an embodiment of the present application. DETAILED DESCRIPTION

[0008] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions of the present application are further elaborated in detail below in conjunction with the drawings and embodiments. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0009] The embodiment of the present application provides a reservoir irrigation area flood prevention early warning method based on big data analysis, which can be executed by a processor of a computer system. The computer system can refer to a server, a laptop, a tablet computer, a desktop computer, a smart TV, a mobile device (such as a mobile phone, a portable video player, a personal digital assistant, a dedicated messaging device, a portable gaming device) and other devices with data processing capabilities.

[0010] Figure 1 A schematic diagram of the implementation process of a reservoir irrigation area flood prevention early warning method based on big data analysis provided in an embodiment of the present application is shown in FIG. Figure 1 As shown, the method includes:

[0011] Step S100: Acquire monitoring data of the reservoir irrigation area for flood forecasting.

[0012] In the embodiment of the present application, the monitoring data of the reservoir irrigation area to be predicted for flood information can be obtained through a variety of ways, including but not limited to:

[0013] 1. Sensor network: Reservoir irrigation areas are usually equipped with a large number of sensors to monitor key parameters such as water level, rainfall, soil moisture, and water flow rate in real time. These sensors send data to the computer system at regular intervals to ensure the real-time and accuracy of the data.

[0014] 2. Remote sensing technology: Satellite remote sensing, drone aerial photography and other technologies are also widely used in the monitoring of reservoir irrigation areas. By analyzing satellite images and high-definition photos taken by drones, computer systems can obtain information such as the topography and vegetation coverage of the irrigation area, which is of great significance for understanding the hydrological process of the irrigation area.

[0015] 3. Historical database: The computer system will also extract relevant data from the existing historical database. These historical data may include rainfall, water level records, flood event records, etc. in the past few years, which are very helpful for analyzing the laws and trends of flood occurrence.

[0016] 4. Manual observation records: In some remote areas or areas with limited technical conditions, manual observation records are still an important way to obtain monitoring data. The computer system will regularly collect these manual observation data and integrate them into a unified database.

[0017] There are many types of monitoring data obtained for reservoir irrigation areas, including but not limited to water level data: recording the water level height of each monitoring point in the reservoir and irrigation area, usually in meters. For example, the water level data of a monitoring point at a certain moment may be "3.5 meters". Rainfall data: recording the rainfall at each rainfall monitoring point in the irrigation area, usually in millimeters. For example, the rainfall data of a monitoring point during a rainfall may be "50 mm". Soil moisture data: obtained through soil moisture sensors, reflecting the moisture content in the soil. For example, the soil moisture data of a monitoring point at a certain moment may be "60%". Water flow velocity data: obtained through equipment such as flow meters, reflecting the speed of water flow. For example, the water flow velocity data of a section of a river may be "2 meters / second". Topographic data: obtained through remote sensing technology, including information such as elevation, slope, and slope direction. For example, the elevation data of a certain area may form an elevation matrix, indicating the height of different locations in the area.

[0018] The acquired raw data usually has different formats and qualities. The computer system preprocesses it to ensure the consistency and availability of the data. For example, duplicate data, invalid data and outliers are removed. For example, if the water level data of a monitoring point suddenly changes to an abnormally high value, it may be caused by a sensor failure. The computer system will mark it as invalid data and delete it. Convert data from different sources and formats into a unified format. For example, convert rainfall data of different units into millimeter units, and convert data of different time resolutions into hourly resolution. For missing data points, the computer system will use interpolation methods (such as linear interpolation, spline interpolation, etc.) to estimate to ensure the integrity of the data. Convert data of different dimensions into the same dimension to facilitate subsequent data processing and model training. For example, normalize data such as water level and rainfall to the range of 0 to 1. The preprocessed data is integrated into a unified database to facilitate subsequent data analysis and model training. Databases are usually managed by relational databases or non-relational databases to support efficient data query and retrieval. At the same time, in order to speed up the data processing and model training process, the computer system may also store some data in an in-memory database or a distributed file system.

[0019] Step S200: Load the reservoir irrigation area monitoring data into the pre-debugging characterization coding network component in the Hongxun prediction neural network for characterization coding to obtain a coded representation of the monitoring data; the pre-debugging characterization coding network component is obtained by pre-debugging the characterization coding network component in the Hongxun prediction neural network without labels based on the pre-debugging reservoir irrigation area monitoring database.

[0020] Prior to step S200, the computer system has completed the pre-debugging (i.e., pre-training) process of the representation encoding network component in the Hongxun prediction neural network. This process is based on a large data set called the pre-debugging reservoir irrigation area monitoring database, which contains a large amount of unlabeled (i.e., unlabeled) reservoir irrigation area monitoring data. These data may include historical water level records, rainfall data, soil moisture data, topographic information, etc., covering a variety of weather conditions and flood events. The purpose of pre-debugging is to allow the representation encoding network component to learn the potential structure and features in the data through unsupervised learning, so that new monitoring data can be effectively encoded. Unsupervised learning is a machine learning method that does not require labeled data, but discovers patterns and laws in the data through the characteristics of the data itself. During the pre-debugging process, the computer system may use an autoencoder, a deep belief network (DBN), or other unsupervised learning algorithms to train the representation encoding network component.

[0021] Take the autoencoder as an example. It is a neural network that learns data representation by reconstructing input data. During the pre-commissioning process, the autoencoder will try to compress the input reservoir irrigation area monitoring data into a low-dimensional encoding representation, and then reconstruct the original data from this encoding representation. By optimizing the reconstruction error, the autoencoder can learn the key features in the data and represent these features in the form of encoding.

[0022] After pre-debugging, the computer system obtained a trained representation encoding network component, which can effectively represent and encode the new reservoir irrigation area monitoring data.

[0023] In step S200, the computer system loads the collected reservoir irrigation area monitoring data into the pre-debugging characterization coding network component for characterization coding. This process includes, for example: before inputting the monitoring data into the characterization coding network component, the computer system may first perform some preprocessing operations on it, such as data cleaning, data conversion, data normalization, etc., to ensure the quality and consistency of the data. The preprocessed monitoring data is input into the characterization coding network component. These data may be a multidimensional array or vector, containing data from multiple monitoring points. Inside the characterization coding network component, the data will undergo a series of transformations and calculations and will eventually be encoded into a low-dimensional vector or array. This encoding process may involve multiple levels of neural networks, each of which will perform different transformations and abstractions on the data. After the encoding process, the computer system obtains a coded representation, which is a low-dimensional and efficient representation of the original monitoring data. This coded representation contains the key features and information in the data and can be used for subsequent Hongxun predictions. The coded representation is an abstraction and generalization of the original monitoring data by the characterization coding network component. It converts high-dimensional raw data into a low-dimensional encoding form by removing redundant information and retaining key features. This conversion not only reduces the dimension and complexity of the data, but also retains important information in the data, making subsequent data processing and analysis more efficient and accurate. In flood forecasting, the representation encoding network component obtained through pre-debugging can learn the potential structure and features in the data, which are of great significance for flood forecasting. After encoding the monitoring data into these feature representations, the computer system can more accurately capture the precursors and trends of flood events, thereby improving the accuracy of the forecast. The original monitoring data usually has high dimensions and complexity, and directly processing and analyzing it may consume a lot of computing resources. After converting the data into a low-dimensional encoding representation through representation encoding, the computational complexity is greatly reduced, allowing the subsequent flood forecasting model to run more efficiently. The pre-debugging process enables the representation encoding network component to learn the data features under different weather conditions and flood events. Therefore, even in the face of new and unseen monitoring data, the encoding representation can effectively extract the key information therein, thereby enhancing the generalization ability of the flood forecasting model. For example, suppose a computer system is encoding the monitoring data of a reservoir irrigation area. The data includes information such as water level, rainfall, soil moisture, etc. at multiple monitoring points. During the pre-commissioning process, the computer system has trained an autoencoder as the representation encoding network component.

[0024] The computer system inputs the monitoring data at a certain moment into the autoencoder. These data may form a multidimensional array, for example:

[0025] [[3.5, 20, 70], i.e., water level, rainfall, and soil moisture at monitoring point 1;

[0026] [3.2, 18, 68], i.e., water level, rainfall, and soil moisture at monitoring point 2;

[0027] …];

[0028] The autoencoder will encode the data in turn. During the encoding process, the autoencoder may first extract the spatial features of the data through the convolutional layer, and then map these features to a low-dimensional encoding space through the fully connected layer. Finally, the computer system obtains an encoded representation, for example:

[0029] [[0.5, -0.2, 0.3], which is the encoding representation of monitoring point 1;

[0030] [0.4, -0.1, 0.2], which is the encoding representation of monitoring point 2;

[0031] …].

[0032] This coded representation is an abstraction and summary of the original monitoring data, which contains the key features and information in the data. In the subsequent Hongxun prediction, the computer system can use these coded representations as input and perform predictive analysis through the Hongxun reasoning component.

[0033] Step S300: Perform Hongxun prediction on the encoded representation of the monitoring data according to the Hongxun reasoning component in the Hongxun prediction neural network to obtain the Hongxun trigger score corresponding to the reservoir irrigation area monitoring data.

[0034] The Hongxun reasoning component is an important part of the Hongxun prediction neural network, which is responsible for making Hongxun predictions based on the encoded representation of the input monitoring data. In the Hongxun prediction neural network, the Hongxun reasoning component is usually a complex neural network model, which may contain multiple levels of neurons and connections to capture nonlinear relationships and patterns in the data.

[0035] The Hongxun reasoning component receives the representative and encoded monitoring data as input. These data exist in the form of encoded representations, which are an abstraction and generalization of the original monitoring data and contain key features and information in the data. Inside the Hongxun reasoning component, the input data will undergo a series of transformations and calculations to extract features that are useful for Hongxun prediction. These features may be statistics, frequency domain characteristics, spatial distribution, etc. of the data, which can reflect the hydrological conditions and potential flood risks of the reservoir irrigation area. The Hongxun reasoning component will use the extracted features to identify patterns in the data, which may be associated with known flood events. By comparing the feature differences between the input data and known flood events, the Hongxun reasoning component can determine whether the current reservoir irrigation area is in a high-risk state for flooding. Based on the results of pattern recognition, the Hongxun reasoning component will output a Hongxun trigger score, which reflects the possibility and severity of flood events in the reservoir irrigation area. The higher the score, the greater the flood risk, and more urgent flood prevention measures need to be taken.

[0036] In step S300, the computer system inputs the monitoring data after characterization encoding into the Hongxun reasoning component to perform Hongxun prediction. The computer system first loads the encoded representation of the monitoring data obtained in step S200 into the memory, ready to be input into the Hongxun reasoning component. At the same time, the computer system also loads the trained Hongxun prediction neural network model, especially the Hongxun reasoning component part. This model may be a deep learning model, such as a convolutional neural network (CNN), a recurrent neural network (RNN), or a combination thereof. After the encoded representation is input into the Hongxun reasoning component, the computer starts the prediction calculation process. This process may involve a large number of matrix operations and nonlinear transformations to calculate the Hongxun trigger score. After the prediction calculation is completed, the computer system outputs the Hongxun trigger score. This score may be a value between 0 and 1, indicating the probability of a flood event; it may also be a classification label, such as "low risk", "medium risk" and "high risk". The Hongxun trigger score is an important result output by the Hongxun prediction neural network, which has a direct indicative effect on judging whether a flood event is about to occur in the reservoir irrigation area.

[0037] For example, suppose that a computer system is predicting floods based on monitoring data from a reservoir irrigation area. In step S200, the system has obtained the coded representations of multiple monitoring points in the reservoir irrigation area, which reflect key information such as water level, rainfall, soil moisture, etc. at each monitoring point. Now, in step S300, the system inputs these coded representations into the flood inference component. The flood inference component may be a convolutional neural network (CNN), which first extracts the spatial features of the input data through the convolution layer, then reduces the dimension and complexity of the data through the pooling layer, and finally outputs the flood trigger score through the fully connected layer. Suppose that the coded representation of a certain monitoring point is: [0.5, -0.2, 0.3, 0.1, -0.4] After this coded representation is input into the CNN, after a series of calculations and transformations, a flood trigger score, such as 0.85, is finally output. This score indicates that the area where the monitoring point is located is more likely to have a flood event, and close attention needs to be paid and corresponding flood prevention measures should be taken.

[0038] Please refer to Figure 2 , is a flow chart of the calibration process of the Hongxun prediction neural network provided in the embodiment of the present application, which specifically includes the following steps:

[0039] Step S10: Obtain the pre-debugging reservoir irrigation area monitoring database, the reservoir irrigation area monitoring data template and the prior flood signal trigger score corresponding to the reservoir irrigation area monitoring data template.

[0040] The pre-commissioning reservoir irrigation area monitoring database is a large data set containing a large amount of historical monitoring data, which covers many aspects of the reservoir irrigation area, such as water level, rainfall, soil moisture, topography, etc. These data are usually collected through sensor networks, remote sensing technology, manual observation and other methods, and are highly diverse and complex.

[0041] The main purpose of pre-commissioning the reservoir irrigation area monitoring database is to provide enough training samples for the representation encoding network component in the Hongxun prediction neural network, so as to learn the potential structure and features in the data through unsupervised learning. These features and structures are crucial for subsequent Hongxun predictions because they can help the neural network better understand and process new monitoring data.

[0042] For example, the pre-commissioning reservoir irrigation area monitoring database may contain water level records of a reservoir irrigation area over the past ten years. These records exist in the form of time series, and each time point corresponds to a water level value. By analyzing this data, the computer system can learn the laws and trends of water level changes, such as seasonal changes and the impact of extreme weather events on water levels.

[0043] The reservoir irrigation area monitoring data template is a set of monitoring data samples with a specific format and structure. They are usually carefully selected from the pre-commissioning reservoir irrigation area monitoring database and are representative and typical. These template data are used for further adjustment and verification of the Hongxun prediction neural network to ensure that the neural network can accurately perform Hongxun predictions on new monitoring data. The reservoir irrigation area monitoring data template may contain data from multiple monitoring points, each of which has corresponding water level, rainfall, soil moisture and other information. These data templates not only contain the actual monitored data values, but may also contain some additional information, such as the geographical location of the monitoring point, topographical features, etc. This information helps the neural network better understand the physical meaning behind the data, thereby improving the accuracy of the prediction. For example, a reservoir irrigation area monitoring data template may contain the following data:

[0044] Monitoring point 1: {"Water level": 3.5 meters,"Rainfall": 20 mm,"Soil moisture": 70%,"Geographic location": [longitude, latitude],"Topography": "Plain"}

[0045] Monitoring point 2:…

[0046] Prior flood signal trigger scores are a set of predetermined score values ​​(i.e., label values) corresponding to the reservoir irrigation area monitoring data template, which represent the possibility and severity of flood events under these specific monitoring data conditions. These score values ​​are usually obtained through historical data analysis, expert evaluation, or simulation experiments, and are highly reliable and accurate.

[0047] In the process of tuning the Hongxun prediction neural network, the prior Hongxun trigger score is used as a true label or reference standard to evaluate the prediction performance of the neural network. By comparing the prediction results of the neural network with the prior Hongxun trigger score, the computer system can calculate the prediction error and adjust and optimize the parameters of the neural network accordingly.

[0048] For example, for the above-mentioned reservoir irrigation area monitoring data template, the prior flood signal trigger score may be a value between 0 and 1, such as 0.85. This score indicates that under the given monitoring data conditions, the possibility of a flood event is high, and close attention needs to be paid and corresponding flood prevention measures should be taken.

[0049] In step S10, the computer system obtains the pre-commissioned reservoir irrigation area monitoring database, the reservoir irrigation area monitoring data template, and the prior flood signal trigger score in a variety of ways. These data may come from different data sources, such as public data from government departments, research results from scientific research institutions, and private data from enterprises (which need to be obtained under the permission of laws and regulations). In order to ensure the quality and reliability of the data, the computer system usually strictly screens, cleans, and verifies these data.

[0050] After obtaining this data, the computer system will use it in subsequent steps to adjust the flood forecasting neural network. Specifically, this data will be used to train the representation encoding network components, verify the prediction performance of the neural network, and optimize the parameter settings of the neural network. Through these steps, the computer system can build an efficient and accurate flood forecasting model to provide strong support for flood prevention work in reservoir irrigation areas.

[0051] Step S20: performing label-free pre-debugging on the representation coding network component according to the reservoir irrigation area monitoring data in the pre-debugging reservoir irrigation area monitoring database to obtain the pre-debugging representation coding network component.

[0052] Step S20 uses a large amount of unlabeled monitoring data in the pre-debugging reservoir irrigation area monitoring database to perform unsupervised learning on the representation coding network component, i.e., unlabeled pre-debugging. This process is designed to enable the representation coding network component to automatically learn useful feature representations from the data, providing a solid foundation for subsequent Hongxun predictions.

[0053] The main purpose of pre-debugging is to enable the representation encoding network component to learn the intrinsic structure and characteristics of the data from a large amount of unlabeled monitoring data through unsupervised learning. These features and structures are crucial for subsequent Hongxun predictions because they can help the neural network better understand and process new monitoring data, thereby improving the accuracy and reliability of the prediction. Under the framework of unsupervised learning, the computer system does not need to rely on labeled data (i.e., known Hongxun trigger scores) for learning. Instead, it uses the characteristics of the data itself to discover patterns and regularities in the data. This learning method enables computer systems to process larger data sets while avoiding the problem of insufficient labeled data or inaccurate labels.

[0054] In step S20, the computer system uses an unsupervised learning algorithm to pre-debug the representation encoding network component. Common unsupervised learning algorithms include autoencoders, deep belief networks (DBNs), generative adversarial networks (GANs), etc. These algorithms can enable the network to learn the feature representation of data in different ways.

[0055] Take the autoencoder as an example. It is a neural network that learns data representation by reconstructing input data. The autoencoder consists of two parts: the encoder and the decoder. The encoder compresses the input data into a low-dimensional encoded representation, while the decoder attempts to reconstruct the original data from this encoded representation. By optimizing the reconstruction error (i.e., the difference between the original data and the reconstructed data), the autoencoder can learn the key features in the data and represent these features in the form of encoding.

[0056] In the implementation process of step S20, the computer system first extracts a large amount of unlabeled monitoring data from the pre-commissioning reservoir irrigation area monitoring database. These data may include various types of information such as water level, rainfall, soil moisture, topography, etc.

[0057] Next, the computer system feeds this data into a representation encoding network component for unsupervised learning. In this process, the representation encoding network component (such as the encoder part of the autoencoder) attempts to compress the input data into a low-dimensional encoding representation. This encoding representation should be able to capture the key features of the data while removing redundant information.

[0058] In order to optimize the quality of the encoded representation, computer systems usually use some optimization algorithms, such as gradient descent, stochastic gradient descent, etc. These algorithms iteratively adjust network parameters to minimize the reconstruction error or other loss functions, thereby gradually improving the effect of the encoded representation.

[0059] After multiple iterations of training, the representation coding network component can learn a stable and effective coding representation. At this point, the computer system saves this pre-debugged representation coding network component as the basis for subsequent Hongxun predictions.

[0060] Pre-debugging allows the representation encoding network components to learn useful feature representations from a large amount of data through unsupervised learning, providing a solid foundation for subsequent Hongxun predictions. These feature representations can capture key information in the data and help the neural network better understand and process new monitoring data. Secondly, pre-debugging improves the generalization ability of Hongxun's prediction neural network. Since a large amount of unlabeled data is used in the pre-debugging process, the neural network can learn more generalized feature representations instead of just relying on limited labeled data. This enables the neural network to make better predictions when faced with new and unseen monitoring data. Finally, pre-debugging also reduces the dependence on labeled data. In practical applications, labeled data is often difficult to obtain or costly. Through pre-debugging, the computer system can learn useful feature representations without labeled data, thereby reducing the dependence on labeled data.

[0061] Step S30: Perform flood prediction on the reservoir irrigation area monitoring data template by pre-debugging the representation coding network component and the flood prediction reasoning component to obtain the predicted flood prediction trigger score.

[0062] In step S30, the computer system first loads the pre-debugging representation encoding network component and Hongxun reasoning component. These two components are trained respectively in the previous steps through unsupervised learning and supervised learning (if applicable). The representation encoding network component is responsible for converting the input monitoring data into low-dimensional, efficient encoding representations, while the Hongxun reasoning component performs Hongxun predictions based on these encoding representations. Next, the computer system inputs the reservoir irrigation area monitoring data templates into the representation encoding network component. These data templates are carefully selected from the pre-debugging reservoir irrigation area monitoring database and are representative and typical. They may contain data from multiple monitoring points, such as water level, rainfall, soil moisture, etc., and these data exist in the form of arrays or vectors.

[0063] For example, a monitoring data template might contain the following data:

[0064] Monitoring point 1: [3.5, 20, 70], i.e., water level (m), rainfall (mm), soil moisture (%);

[0065] Monitoring point 2: [3.2, 18, 68]

[0066] …

[0067] After the representation encoding network component receives this data, it will encode it and convert it into a low-dimensional encoding representation. These encoding representations may be a series of values ​​that capture the key features and information in the original data.

[0068] The computer system then inputs these encoded representations into the flood information component. The flood information component may be a multi-layer perceptron (MLP), convolutional neural network (CNN), or other types of neural network models. It performs complex calculations and analyses based on the encoded representations to predict the likelihood and severity of flood events.

[0069] The output of the flood information reasoning component is a predicted flood information trigger score, which indicates the risk level of a flood event under the current monitoring data conditions. This score may be a value between 0 and 1, where 0 indicates no risk and 1 indicates extremely high risk. The higher the score, the greater the flood risk and the more urgent flood prevention measures need to be taken.

[0070] In step S30, the implementation of the representation encoding network component and the Hongxun reasoning component may vary depending on the specific application scenario and data characteristics. However, no matter which model is used, the core idea is to make predictions by learning the feature representation of the data.

[0071] Take the convolutional neural network (CNN) as an example, which is usually used to process data with spatial structure (such as images). In Hongxun prediction, if the monitoring data contains topographic information (such as satellite images), CNN can be used as the Hongxun reasoning component. CNN extracts the spatial features of the data through structures such as convolutional layers, pooling layers, and fully connected layers, and makes predictions based on these features. On the other hand, if the monitoring data is mainly time series data (such as water level, rainfall records), recurrent neural network (RNN) or its variants (such as LSTM, GRU) can be used as Hongxun reasoning components. RNN is able to handle temporal dependencies in sequence data and is very effective in predicting future trends in time series data.

[0072] In the Hongxun reasoning component, the calculation method for predicting the Hongxun trigger score may involve complex mathematical operations and the forward propagation process of the neural network. These calculations are usually completed efficiently under the hardware acceleration of the computer system.

[0073] The predicted flood signal trigger score output in step S30 is an important basis for evaluating the flood risk of the reservoir irrigation area. The computer system can determine whether it is necessary to issue a flood warning, deploy flood control resources or take other emergency measures based on this score.

[0074] For example, if the predicted flood signal trigger score of a certain monitoring area exceeds the preset threshold (such as 0.8), the computer system can automatically trigger the flood prevention warning mechanism and issue warning information to relevant departments and the public. At the same time, it can also deploy flood prevention materials and personnel according to the score to ensure the smooth progress of flood prevention work.

[0075] Step S40: Determine the prediction error according to the predicted flood signal trigger score and the prior flood signal trigger score.

[0076] Step S40 determines the prediction error based on the predicted flood signal trigger score and the prior flood signal trigger score. The prediction error refers to the difference between the predicted value and the actual value. In step S40, the predicted value refers to the predicted flood signal trigger score output by the flood signal prediction neural network, and the actual value refers to the prior flood signal trigger score, that is, the flood risk score that actually occurred under the same monitoring data conditions based on historical data or expert judgment. The size of the prediction error reflects the accuracy of the flood signal prediction neural network. The smaller the error, the closer the prediction result of the neural network is to the actual situation, and the better the prediction performance; conversely, the larger the error, the larger the deviation between the prediction result of the neural network and the actual situation, and further optimization is needed.

[0077] In step S40, the computer system calculates the prediction error by comparing the predicted flood trigger score with the prior flood trigger score. The specific calculation method may vary depending on the application scenario and data characteristics, but common error metrics include mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE), etc. In addition to MSE, mean absolute error (MAE) and root mean square error (RMSE) are also commonly used error metrics. MAE quantifies the prediction error by calculating the average of the absolute values ​​of the difference between the predicted value and the actual value of each sample.

[0078] For example, suppose there is a set of reservoir irrigation area monitoring data templates, each template contains water level, rainfall, soil moisture and other data of multiple monitoring points, and each template corresponds to a prior flood trigger score. After calculation by the flood prediction neural network, the predicted flood trigger score of each template is obtained. Taking one of the templates as an example, suppose its prior flood trigger score is 0.7 (indicating medium to high risk), and the predicted flood trigger score output by the flood prediction neural network is 0.65. At this point, MSE can be used to calculate the prediction error: MSE=(0.7-0.65) 2 = 0.0025. This value indicates that the difference between the predicted value and the actual value is very small, indicating that the prediction performance of the neural network is good.

[0079] Prediction error plays a vital role in the flood prevention and warning method for reservoir irrigation areas based on big data analysis. First, it is an important indicator for evaluating the prediction performance of neural networks. By comparing the prediction errors of different models, the model with the best performance can be selected for deployment. Secondly, the prediction error can also guide the tuning process of the neural network. In step S50, the computer system adjusts the parameters of the Hongxun prediction neural network according to the prediction error to reduce the prediction error and improve the prediction accuracy. This process usually involves the application of optimization algorithms such as gradient descent. Finally, the prediction error can also be used to monitor the stability of the neural network. If the prediction error changes significantly in a short period of time, it may mean that the neural network has problems such as overfitting or underfitting, and timely measures need to be taken to adjust it.

[0080] Step S50: Adjusting the component parameters of the pre-debugging characterization encoding network component and / or the Hongxun reasoning component based on the prediction error.

[0081] Step S50 adjusts the component parameters of the pre-debugging characterization encoding network component and / or the Hongxun reasoning component in the Hongxun prediction neural network according to the prediction error.

[0082] The purpose of the calibration is to optimize the performance of the neural network model by adjusting its parameters so that it can more accurately predict the flood trigger score of the reservoir irrigation area. The prediction error is an important indicator to measure the performance of the model, which reflects the difference between the model's predicted value and the actual value. By adjusting the model parameters, this difference can be reduced and the prediction accuracy of the model can be improved.

[0083] In step S40, the computer system has calculated the prediction error based on the predicted flood signal trigger score and the prior flood signal trigger score. These error values ​​provide direct feedback on the performance of the model, indicating where the model is deficient. The purpose of the calibration is to adjust these deficiencies so that the model can better adapt to the data characteristics and improve the prediction accuracy.

[0084] Methods for adjusting neural network model parameters include, for example, optimization algorithms such as gradient descent, stochastic gradient descent, and mini-batch gradient descent. These methods optimize model performance by iteratively adjusting model parameters to minimize the loss function (i.e., prediction error).

[0085] Taking the gradient descent method as an example, its basic principle is to iteratively update the gradient direction of the model parameters along the loss function in the hope of finding the parameter value that minimizes the loss function. In each iteration, the gradient of the loss function under the current parameters is calculated, and the model parameters are updated according to the learning rate (a hyperparameter that controls the parameter update step size) and the gradient direction.

[0086] In step S50, the calibration process of the computer system can be summarized as the following steps:

[0087] 1. Initialization parameters: Before starting to tune, you need to initialize the parameters of the neural network model, including weights and biases, etc. These parameters are usually initialized randomly or based on prior knowledge.

[0088] 2. Forward propagation: Input the reservoir irrigation area monitoring data template into the neural network model, and obtain the predicted flood trigger score through forward propagation calculation.

[0089] 3. Calculate the loss: Calculate the loss function value (i.e., prediction error) based on the predicted Hongxun trigger score and the prior Hongxun trigger score.

[0090] 4. Back propagation: The gradient of the loss function to the model parameters is calculated through the back propagation algorithm.

[0091] 5. Parameter update: Update the model parameters according to the calculated gradient and learning rate.

[0092] 6. Iterative optimization: Repeat steps 2 to 5 until the preset number of iterations is reached or the loss function value converges.

[0093] After the adjustment in step S50, the performance of the neural network model will usually be improved, the prediction error will be reduced, and the prediction accuracy will be improved. These optimized model parameters can be saved and used for subsequent Hongxun prediction tasks.

[0094] As an implementation mode, step S20, performing label-free pre-debugging on the representation coding network component according to the reservoir irrigation area monitoring data in the pre-debugging reservoir irrigation area monitoring database to obtain the pre-debugging representation coding network component, includes:

[0095] Step S21: acquiring the real reservoir irrigation area monitoring data, the early warning reservoir irrigation area monitoring data and the normal reservoir irrigation area monitoring data in the pre-debugging reservoir irrigation area monitoring database; the real reservoir irrigation area monitoring data represents the reservoir irrigation area monitoring data corresponding to the early warning state, the early warning reservoir irrigation area monitoring data represents the reservoir irrigation area monitoring data obtained after the reservoir irrigation area monitoring data is transformed into the early warning state, and the normal reservoir irrigation area monitoring data represents the reservoir irrigation area monitoring data other than the real reservoir irrigation area monitoring data and the early warning reservoir irrigation area monitoring data in the pre-debugging reservoir irrigation area monitoring database;

[0096] Step S22: Perform label-free pre-debugging on the representation coding network component according to the real reservoir irrigation area monitoring data, the early warning reservoir irrigation area monitoring data and the normal reservoir irrigation area monitoring data to obtain the pre-debugging representation coding network component.

[0097] In step S21, the computer system obtains three types of monitoring data from the pre-debugging reservoir irrigation area monitoring database: real reservoir irrigation area monitoring data, early warning reservoir irrigation area monitoring data and normal reservoir irrigation area monitoring data.

[0098] Authentic reservoir irrigation area monitoring data: This type of data directly reflects the actual situation of the reservoir irrigation area under the warning state. For example, when the reservoir water level reaches or exceeds the warning line, the relevant monitoring data (such as water level, rainfall, soil moisture, etc.) are regarded as authentic reservoir irrigation area monitoring data. These data are essential for training the representation encoding network component to identify the warning state.

[0099] Early warning reservoir irrigation area monitoring data: This type of data is obtained by transforming normal reservoir irrigation area monitoring data into early warning states. The early warning state transformation may include simulating scenarios such as rising water levels and increased rainfall to simulate conditions that may trigger early warnings. In this way, the computer system can generate more diverse early warning state data, which helps to enhance the generalization ability of the representation encoding network components.

[0100] Normal reservoir irrigation area monitoring data: This type of data represents the monitoring situation of the reservoir irrigation area under normal operating conditions. As comparative data, they help characterize the coding network components to distinguish between warning status and normal status and improve the accuracy of prediction.

[0101] For example, assuming that the pre-commissioning reservoir irrigation area monitoring database contains monitoring data from the past year, the computer system can extract records of water levels exceeding the warning line as real reservoir irrigation area monitoring data; at the same time, by randomly selecting some monitoring data in normal state and artificially adjusting their water level values ​​to simulate the warning state, the warning reservoir irrigation area monitoring data is generated; the normal data that is not selected for the warning state transformation is used as the normal reservoir irrigation area monitoring data.

[0102] In step S22, the computer system uses the real reservoir irrigation area monitoring data, the early warning reservoir irrigation area monitoring data, and the normal reservoir irrigation area monitoring data obtained in step S21 to perform unlabeled pre-debugging on the representation encoding network component. This process is intended to enable the representation encoding network component to automatically learn the potential structure and features in the data through unsupervised learning.

[0103] The implementation of unlabeled pre-debugging varies depending on the specific structure of the representation encoding network component, for example, including: before the monitoring data is input into the representation encoding network component, the computer system first pre-processes it. This includes steps such as data cleaning (such as removing outliers, filling missing values, etc.), data standardization (such as normalization, standardization, etc.) and data enhancement (such as data enhancement technology to generate more diverse training samples). According to the design of the representation encoding network component, the computer system initializes its network structure and parameters. This may include setting the number of layers of the network, the number of neurons in each layer, the activation function, etc. The pre-processed monitoring data is input into the representation encoding network component, and the encoding representation is obtained by forward propagation calculation. The encoding representation is a low-dimensional and efficient representation of the data, which retains the key information in the data while removing redundancy and noise. For some types of representation encoding network components (such as autoencoders), a reconstruction process is also included. In this process, the encoding representation is used as input and the reconstructed data is calculated by back propagation. By optimizing the reconstruction error (that is, the difference between the original data and the reconstructed data), the representation encoding network component can learn a more effective encoding representation. In the framework of unsupervised learning, loss functions are usually used to measure the quality of the encoding representation. For autoencoders, commonly used loss functions include mean square error (MSE), cross entropy loss, etc. The computer system optimizes the parameters of the representation encoding network component by minimizing the loss function. The parameters of the representation encoding network component are updated according to the gradient of the loss function using an optimization algorithm (such as gradient descent, Adam optimizer, etc.). This process is usually completed through multiple iterations until the loss function converges or reaches a preset number of iterations. After completing the unlabeled pre-debugging, the computer system will evaluate the representation encoding network component. This can be done by calculating the clustering effect of the encoding representation, visual analysis, etc. Based on the evaluation results, the computer system can select the representation encoding network component with the best performance as the pre-debugging representation encoding network component.

[0104] Taking the autoencoder as an example, it is assumed that the representation encoding network component is a multi-layer perceptron (MLP) structure. In step S22, the computer system first preprocesses the real reservoir irrigation area monitoring data, the warning reservoir irrigation area monitoring data and the normal reservoir irrigation area monitoring data; then inputs the preprocessed data into the autoencoder; obtains the encoding representation through forward propagation calculation; then calculates the reconstruction error as the loss function; uses the Adam optimizer to update the weights and biases of the autoencoder according to the gradient of the loss function; finally, optimizes the parameters of the autoencoder through multiple iterations. After completing the unlabeled pre-debugging, the computer system can select the autoencoder with the smallest reconstruction error as the pre-debugging representation encoding network component.

[0105] Through the implementation of steps S21 and S22, the computer system can use the unlabeled monitoring data in the pre-debugging reservoir irrigation area monitoring database to perform unsupervised learning on the representation coding network component, thereby obtaining a pre-debugging representation coding network component that can effectively extract data features.

[0106] As an implementation mode, step S22, based on the real reservoir irrigation area monitoring data, the early warning reservoir irrigation area monitoring data and the normal reservoir irrigation area monitoring data, the representation coding network component is pre-debugged without labels to obtain the pre-debugged representation coding network component, including:

[0107] Step S221: obtaining the noisy original true reservoir irrigation area monitoring data corresponding to the original true reservoir irrigation area monitoring data and the noisy early warning reservoir irrigation area monitoring data corresponding to the early warning reservoir irrigation area monitoring data; the noisy original true reservoir irrigation area monitoring data represents the reservoir irrigation area monitoring data obtained by transforming the early warning status and / or early warning level in the original true reservoir irrigation area monitoring data but not changing the basic information; the noisy early warning reservoir irrigation area monitoring data represents the reservoir irrigation area monitoring data obtained by transforming the basic information in the early warning reservoir irrigation area monitoring data but not changing the early warning status and early warning level;

[0108] Step S222: characterizing and encoding the real reservoir irrigation area monitoring data, the noisy real reservoir irrigation area monitoring data and the normal reservoir irrigation area monitoring data respectively according to the characterizing and encoding network component, and obtaining the real monitoring data encoding representation corresponding to the real reservoir irrigation area monitoring data, the noisy real monitoring data encoding representation corresponding to the noisy real reservoir irrigation area monitoring data and the normal monitoring data encoding representation corresponding to the normal reservoir irrigation area monitoring data;

[0109] Step S223: Calculate according to the coded representation of the true monitoring data, the coded representation of the noisy true monitoring data and the coded representation of the normal monitoring data to obtain a first error value;

[0110] Step S224: characterize and encode the early warning reservoir irrigation area monitoring data and the noisy early warning reservoir irrigation area monitoring data respectively according to the characterization encoding network component, and obtain the early warning monitoring data encoding representation corresponding to the early warning reservoir irrigation area monitoring data and the noisy early warning monitoring data encoding representation corresponding to the noisy early warning reservoir irrigation area monitoring data;

[0111] Step S225: Calculate according to the early warning monitoring data coding representation, the noisy early warning monitoring data coding representation and the normal monitoring data coding representation to obtain a second error value;

[0112] Step S226: calibrate the component parameters of the characterizing coding network component according to the first error value and the second error value to obtain a pre-debugging characterizing coding network component.

[0113] In step S221, the computer system first performs noise processing on the real reservoir irrigation area monitoring data and the early warning reservoir irrigation area monitoring data to generate noisy real reservoir irrigation area monitoring data and noisy early warning reservoir irrigation area monitoring data. The purpose of this step is to increase the diversity of the data by introducing noise, so as to help the representation encoding network component learn more robust feature representation.

[0114] Noise-added Benzhen Reservoir Irrigation Area Monitoring Data: This type of data is obtained by transforming the warning status and / or warning level in the Benzhen Reservoir Irrigation Area Monitoring Data. For example, if the Benzhen data represents the flood warning status, the noise-added data may increase or decrease the warning level by one level, but does not change the basic information in the data (such as water level, rainfall, etc.).

[0115] Noise-added warning reservoir irrigation area monitoring data: This type of data is obtained by transforming the basic information in the warning reservoir irrigation area monitoring data, while keeping the warning state and warning level unchanged. For example, the noise-added data can be generated by randomly disturbing the water level value, rainfall value, etc.

[0116] For example, suppose there is a record in the monitoring data of the Benzhen Reservoir Irrigation District: {"water level": 3.5 meters, "rainfall": 20 millimeters, "warning status": "high"}. After noise processing, the noisy monitoring data of the Benzhen Reservoir Irrigation District may be: {"water level": 3.5 meters, "rainfall": 20 millimeters, "warning status": "extremely high"} (the warning level is increased), or {"water level": 3.6 meters, "rainfall": 22 millimeters, "warning status": "high"} (the basic information is slightly disturbed).

[0117] In step S222, the computer system uses the representation coding network component to perform representation coding on the real reservoir irrigation area monitoring data, the noisy real reservoir irrigation area monitoring data and the normal reservoir irrigation area monitoring data to obtain corresponding coding representations. The representation coding network component may be a multi-layer perceptron (MLP), a convolutional neural network (CNN) or other types of neural networks.

[0118] Encoded representation of authentic monitoring data: The representation encoding network component converts the authentic reservoir irrigation area monitoring data into a low-dimensional encoding vector that contains the key feature information in the data.

[0119] Noised original monitoring data coding representation: Similarly, the noisy original reservoir irrigation area monitoring data is also converted into a coding vector, but due to the noise addition process, this coding vector may be different from the coding representation of the original monitoring data.

[0120] Coding representation of normal monitoring data: Normal reservoir irrigation area monitoring data are also converted into coding vectors as comparative data, which helps to characterize the coding network components to distinguish between warning status and normal status.

[0121] For example, suppose the representation encoding network component is an MLP that accepts an input vector containing features such as water level and rainfall, and outputs a low-dimensional encoding vector. For the real data {"water level": 3.5 meters, "rainfall": 20 mm, "warning status": "high"}, its encoding representation may be a vector of the shape [0.5, -0.2, 0.3].

[0122] In step S223, the computer system calculates the error between the authentic monitoring data encoding representation, the noisy authentic monitoring data encoding representation and the normal monitoring data encoding representation to obtain a first error value. This error value reflects the difference and robustness of the characterization coding network component when processing authentic data and normal data.

[0123] Error calculation: Various distance metrics or similarity metrics can be used to calculate the difference between the encoded representations, such as Euclidean distance, cosine similarity, etc. The first error value may be a weighted sum or average of these metrics.

[0124] For example, assuming that Euclidean distance is used as the error metric, the true monitoring data is encoded as [0.5, -0.2, 0.3], the noisy true monitoring data is encoded as [0.6, -0.1, 0.2], and the normal monitoring data is encoded as [0.4, -0.3, 0.1]. Then the first error value can be calculated as:

[0125] .

[0126] In step S224, the computer system characterizes and encodes the early warning reservoir irrigation area monitoring data and the noisy early warning reservoir irrigation area monitoring data to obtain the early warning monitoring data encoding representation and the noisy early warning monitoring data encoding representation. This step is similar to step S222, but focuses on the data in the early warning state.

[0127] Encoded representation of early warning monitoring data: The representation encoding network component converts the early warning reservoir irrigation area monitoring data into a coded vector.

[0128] Encoded representation of noisy early warning monitoring data: The noisy early warning reservoir irrigation area monitoring data are also converted into encoding vectors to evaluate the stability of the characterization encoding network components when processing noisy early warning data.

[0129] In step S225, the computer system calculates the error between the early warning monitoring data encoding representation, the noisy early warning monitoring data encoding representation and the normal monitoring data encoding representation to obtain a second error value. This error value reflects the difference and robustness of the characterization coding network component when processing early warning data and normal data.

[0130] Similar to step S223, various distance metrics or similarity metrics can be used to calculate the difference between the coded representations, and the second error value can be a weighted sum or average of these metric values. For example, using Euclidean distance as the error metric, the early warning monitoring data is coded as [0.7, 0.0, 0.4], the noisy early warning monitoring data is coded as [0.8, 0.1,0.3], and the normal monitoring data is coded as [0.4, -0.3, 0.1]. Then the second error value can be calculated as: .

[0131] In step S226, the computer system adjusts the parameters of the encoding network component according to the first error value and the second error value. This step is the core of the optimization process, which aims to reduce the error value and improve the performance of the encoding network component by adjusting the network parameters.

[0132] Various optimization algorithms can be used to update network parameters, such as gradient descent method, Adam optimizer, etc. These algorithms guide the update direction of parameters according to the gradient of the error value. Parameter adjustment is usually an iterative process, and steps S222 to S225 need to be repeated many times until the error value converges or reaches a preset number of iterations.

[0133] Through the implementation of steps S221 to S226, the computer system can use noise processing and error calculation to enhance the robustness and generalization ability of the representation coding network component. Noise processing increases the diversity of data by introducing noise, which helps the network learn more robust feature representations; while error calculation provides direct feedback on network performance and guides the optimization direction of parameters. Finally, after multiple iterative adjustments, the computer system can obtain a pre-debugging representation coding network component with superior performance, providing a solid foundation for subsequent Hongxun predictions.

[0134] As an implementation method, the characterization coding network component includes a first characterization coding branch component and a second characterization coding branch component, and the first characterization coding branch component and the second characterization coding branch component have the same composition architecture, but have different component parameters. Based on this, step S222, according to the characterization coding network component, the true reservoir irrigation area monitoring data, the noisy true reservoir irrigation area monitoring data and the normal reservoir irrigation area monitoring data are characterized and encoded respectively to obtain the true monitoring data encoding representation corresponding to the true reservoir irrigation area monitoring data, the noisy true monitoring data encoding representation corresponding to the noisy true reservoir irrigation area monitoring data and the normal monitoring data encoding representation corresponding to the normal reservoir irrigation area monitoring data, including:

[0135] Step S2221: characterize and encode the real reservoir irrigation area monitoring data according to the first characterization and encoding branch component to obtain the real monitoring data encoding representation;

[0136] Step S2222: characterize and encode the noisy original reservoir irrigation area monitoring data according to the second characterization encoding branch component to obtain a coded representation of the noisy original monitoring data;

[0137] Step S2223: Characterize and encode the normal reservoir irrigation area monitoring data according to the first characterization coding branch component or the second characterization coding branch component to obtain a coded representation of the normal monitoring data.

[0138] Step S222 uses the representation coding network component to characterize and encode different types of monitoring data to extract key features and information from the data. In this step, the representation coding network component is designed to include two branch components, a first representation coding branch component and a second representation coding branch component. Although the two branch components have the same composition architecture, they have different component parameters. This design aims to improve the representation ability and robustness of the model by increasing its diversity and complexity.

[0139] In step S2221, the computer system uses the first representation coding branch component to characterize and encode the real reservoir irrigation area monitoring data to obtain the real monitoring data encoding representation. The real reservoir irrigation area monitoring data refers to the monitoring data that directly reflects the actual situation of the reservoir irrigation area under the early warning state, which is crucial for training the representation coding network component to identify the early warning state.

[0140] First representation encoding branch component: This is a complex neural network structure that may contain multiple layers, such as input layer, hidden layer, and output layer. Each layer contains a certain number of neurons, which are connected by weights and biases. In the first representation encoding branch component, these weights and biases are initialized to different values ​​to distinguish them from the second representation encoding branch component.

[0141] Representation encoding process: When the monitoring data of the Benzhen reservoir irrigation area is input into the first representation encoding branch component, it will undergo a series of nonlinear transformations and linear transformations and finally be converted into a low-dimensional encoding vector. This encoding vector contains the key features and information in the data and is the basis for subsequent Hongxun predictions.

[0142] For example, suppose the real reservoir irrigation area monitoring data contains features such as water level, rainfall, soil moisture, etc. These features are input into the first representation encoding branch component in the form of vectors. For example, a real data sample may be: [3.5, 20, 70], representing water level (meters), rainfall (mm), and soil moisture (%). After being processed by the first representation encoding branch component, this sample may be encoded into a low-dimensional vector of the form [0.5, -0.2, 0.3], which is the real monitoring data encoding representation.

[0143] In step S2222, the computer system uses the second representation coding branch component to characterize and encode the noisy original reservoir irrigation area monitoring data to obtain a coding representation of the noisy original monitoring data. The noisy original reservoir irrigation area monitoring data is the result of noise processing on the original data, which aims to increase the diversity of the data by introducing noise, and help the representation coding network component learn a more robust feature representation.

[0144] Second representation encoding branch component: Similar to the first representation encoding branch component, the second representation encoding branch component is also a complex neural network structure, containing multiple layers and neurons. However, its weights and biases are different from those of the first representation encoding branch component, which enables the two branch components to produce different encoding representations when processing the same data.

[0145] Representation encoding process: After the noise-added original reservoir irrigation area monitoring data is input into the second representation encoding branch component, it will undergo a nonlinear transformation and linear transformation process similar to the first branch, and finally be converted into a low-dimensional encoding vector. Due to the influence of the noise addition process, this encoding vector may be different from the encoding representation of the original monitoring data, but it still contains the key features and information in the data.

[0146] For example, suppose that the noisy original reservoir irrigation area monitoring data is obtained by slightly perturbing the original data, such as adjusting the water level value from 3.5 meters to 3.6 meters. After the noisy data sample [3.6, 20, 70] is input into the second representation encoding branch component, it may be encoded into a low-dimensional vector of the form [0.6, -0.1, 0.2], that is, the noisy original monitoring data encoding representation.

[0147] In step S2223, the computer system may select the first characterization coding branch component or the second characterization coding branch component to characterize and code the normal reservoir irrigation area monitoring data to obtain a coded representation of the normal monitoring data. The normal reservoir irrigation area monitoring data represents the monitoring situation of the reservoir irrigation area under normal operating conditions. As comparison data, they help the characterization coding network component distinguish between the warning state and the normal state.

[0148] Branch component selection: In practical applications, you can choose to use the first characterization encoding branch component or the second characterization encoding branch component to process normal reservoir irrigation area monitoring data. This choice may be based on experience, experimental results, or randomness. Regardless of which branch component is selected, its goal is to extract key features and information from normal data.

[0149] Characterization encoding process: After the normal reservoir irrigation area monitoring data is input into the selected branch component, it will go through a nonlinear transformation and linear transformation process similar to the previous steps, and finally be converted into a low-dimensional encoding vector. This encoding vector represents the data characteristics under normal conditions and is an important basis for distinguishing between warning conditions and normal conditions in the future.

[0150] For example, suppose the first representation encoding branch component is selected to process normal reservoir irrigation area monitoring data. After a normal data sample [3.0, 15, 65] is input into the branch component, it may be encoded into a low-dimensional vector of the form [0.4, -0.3, 0.1], which is the encoding representation of normal monitoring data.

[0151] Through the implementation of steps S2221 to S2223, the computer system can use the first characterization coding branch component and the second characterization coding branch component to characterize and encode different types of monitoring data to obtain corresponding coding representations. These coding representations are the basis for subsequent Hongxun predictions. They contain key features and information in the data and help improve the accuracy and reliability of the predictions. At the same time, by introducing two branch components with different parameters, the diversity and complexity of the model are increased, which helps to improve its characterization ability and robustness.

[0152] As an implementation mode, step S2221 to step S2222, characterize and encode the real reservoir irrigation area monitoring data according to the first characterization coding branch component to obtain the real monitoring data encoding representation; characterize and encode the noisy real reservoir irrigation area monitoring data according to the second characterization coding branch component to obtain the noisy real monitoring data encoding representation, including:

[0153] Step S222a: filtering the real reservoir irrigation area monitoring data according to the first filter in the first representation coding branch component to obtain a corresponding real reservoir irrigation area monitoring data vector set;

[0154] Step S222b: performing representation encoding on the true reservoir irrigation area monitoring data vector set according to the first representation encoding unit in the first representation encoding branch component to obtain a true monitoring data encoding representation;

[0155] Step S222c: filtering the noisy original reservoir irrigation area monitoring data according to the second filter in the second representation coding branch component to obtain a corresponding noisy original reservoir irrigation area monitoring data vector set;

[0156] Step S222d: Characterize and encode the noisy true reservoir irrigation area monitoring data vector set according to the second characterization encoding unit in the second characterization encoding branch component to obtain a coded representation of the noisy true monitoring data.

[0157] Steps S2221 to S2222 use the first characterization coding branch component and the second characterization coding branch component to characterize and code the real reservoir irrigation area monitoring data and the noisy real reservoir irrigation area monitoring data respectively. In this process, each branch component includes two main parts, a filter and a characterization coding unit, which work together to extract key features from the data and generate a coded representation.

[0158] In step S222a, the computer system uses the first filter in the first characterization coding branch component to filter the real reservoir irrigation area monitoring data. Filtering is a common technique in signal processing, which aims to extract useful information or remove unwanted noise from the original signal. In this scenario, the purpose of the filter is to convert the real reservoir irrigation area monitoring data into a series of vector sets, which can be regarded as feature maps of the data, which capture the spatial structure or temporal relationship in the data.

[0159] The first filter is a pre-designed mathematical function or neural network layer that can perform a specific transformation on the input data. In the first representation encoding branch component, the first filter may be a convolutional layer (if processing image or grid data) or a recurrent layer (if processing time series data). Its parameters (such as the weight of the convolution kernel) are learned during the model training process.

[0160] When the real reservoir irrigation area monitoring data is input into the first filter, the filter performs a convolution operation (for image data) or a loop operation (for time series data) on it to generate a series of feature vectors. These vector sets capture local features in the data, such as edges, textures, or patterns in time series.

[0161] For example, suppose that the real reservoir irrigation area monitoring data contains time series data such as water level and rainfall at multiple monitoring points. The first filter may be a one-dimensional convolution layer with a convolution kernel size of 3, a step size of 1, and a padding of 0. When this convolution layer is applied to the real data, it slides along the time axis and performs a weighted sum of the data in each local window to generate a series of feature vectors. For example, for the input data [x_1, x_2, x_3, x_4, x_5], the convolution operation may produce the output [y_1, y_2, y_3, y_4], where each y_i is the weighted sum of the data in the corresponding local window.

[0162] In step S222b, the computer system uses the first representation encoding unit in the first representation encoding branch component to represent and encode the true reservoir irrigation area monitoring data vector set obtained in step S222a to generate a true monitoring data encoding representation. Representation encoding is an important concept in machine learning, which aims to convert raw data into a more compact and meaningful representation that can retain key information in the data and remove redundancy.

[0163] The first representation encoding unit is a complex neural network structure, which may contain multiple fully connected layers, activation functions, and pooling layers. In the first representation encoding branch component, the first representation encoding unit is responsible for converting the filtered feature vector into a low-dimensional encoding representation.

[0164] When the real reservoir irrigation area monitoring data vector set is input into the first representation encoding unit, it will undergo a series of nonlinear transformations and linear transformations, which are designed to extract deep features in the data and generate encoding representations. These transformations may include matrix multiplication of the fully connected layer, nonlinear activation of the activation function (such as ReLU, Sigmoid, etc.), and downsampling operations of the pooling layer.

[0165] For example, assume that the first representation encoding unit includes a fully connected layer and a ReLU activation function. When the feature vector set [v_1, v_2, ..., v_n] obtained in step S222a is input into this unit, the fully connected layer will perform a linear transformation on it to generate a new vector w=Wv+b, where W is the weight matrix, v is the stacked form of the feature vector set, and b is the bias vector. The ReLU activation function then applies nonlinear activation to each element in w to generate the final encoding representation e=ReLU(w). This encoding representation e is a low-dimensional vector that captures the key features of the authentic reservoir irrigation area monitoring data.

[0166] In step S222c, the computer system uses a similar process as step S222a, but this time the noise-added real reservoir irrigation area monitoring data is filtered. The noise-added real data is the result of noise-added real data, which aims to increase the diversity of the data and improve the robustness of the model by introducing noise.

[0167] The second filter is similar to the first filter. The second filter is also a pre-designed mathematical function or neural network layer. However, since it processes the noisy real data, its parameters may be different from the first filter. This difference helps the second representation encoding branch component learn a different feature representation from the first branch.

[0168] When the noisy original reservoir irrigation area monitoring data is input into the second filter, it will undergo a convolution or loop operation similar to the first filter to generate a series of feature vector sets. These feature vector sets capture the local features in the noisy data, and due to the influence of the noise processing, they may be different from the feature vectors of the original data.

[0169] For example, suppose the noisy original reservoir irrigation area monitoring data is obtained by adding Gaussian noise to the original data. The second filter may be a convolutional layer with the same structure as the first filter but different parameters. When this convolutional layer is applied to the noisy data, it generates a series of new feature vector sets that contain both the characteristics of the original data and the influence of the noise.

[0170] In step S222d, the computer system uses the second representation coding unit in the second representation coding branch component to characterize and code the noisy true reservoir irrigation area monitoring data vector set obtained in step S222c to generate a noisy true monitoring data coding representation. This process is similar to step S222b, but uses the second representation coding unit, which may have different parameters and structures from the first representation coding unit.

[0171] The second representation encoding unit is a neural network structure similar to the first representation encoding unit but with different parameters and structure. It is responsible for converting the filtered noisy true data vector set into a low-dimensional encoding representation.

[0172] When the noisy original reservoir irrigation area monitoring data vector set is input into the second representation coding unit, it will undergo a series of nonlinear transformations and linear transformations, which are designed to extract deep features in the noisy data and generate a coded representation. Due to the influence of the noise processing and the differences in the second representation coding unit, the generated coded representation may be different from the coded representation of the original data.

[0173] For example, assume that the second representation coding unit also includes a fully connected layer and a ReLU activation function, but its weight matrix and bias vector are different from those of the first representation coding unit. When the feature vector set obtained in step S222c is input into this unit, it generates a new encoding representation e'=ReLU(W'v'+b'), where W' and b' are the weight matrix and bias vector of the second representation coding unit, respectively, and v' is the stacked form of the feature vector set of the noisy original data. This encoding representation e' captures the key features of the noisy original data and reflects the impact of noise on data representation.

[0174] Through the implementation of steps S222a to S222d, the computer system can use the first characterization coding branch component and the second characterization coding branch component to filter and characterize the real reservoir irrigation area monitoring data and the noisy real reservoir irrigation area monitoring data, respectively, so as to generate corresponding coded representations. These coded representations are the basis for subsequent Hongxun predictions. They contain key features and information in the data, which helps to improve the accuracy and reliability of the predictions. At the same time, by introducing noise processing and different branch component parameters, the diversity and robustness of the model are increased, which helps to improve its performance in complex environments.

[0175] As an implementation manner, step S223, performing calculation according to the coded representation of the true monitoring data, the coded representation of the noisy true monitoring data, and the coded representation of the normal monitoring data to obtain a first error value, includes:

[0176] Step S2231: Determine a first error value based on a first commonality metric score between the authentic monitoring data coded representation and the noisy authentic monitoring data coded representation, and a second commonality metric score between the authentic monitoring data coded representation and the normal monitoring data coded representation.

[0177] In step S2231, the computer system first calculates a first commonality metric score between the coded representation of the true monitoring data and the coded representation of the noisy true monitoring data, and then calculates a second commonality metric score between the coded representation of the true monitoring data and the coded representation of the normal monitoring data. The two scores reflect the similarity or distance between the true data and the noisy true data, and the true data and the normal data in the coding space. Based on the two scores, the computer system can further calculate a first error value for evaluating the performance of the characterizing coding network component.

[0178] The first commonality metric score is used to quantify the similarity between the encoded representation of the true monitoring data and the encoded representation of the noisy true monitoring data. Since the noisy true data is obtained by slightly perturbing the true data, in theory, the encoded representations of the two should have a high similarity. The commonality metric score can be calculated by a variety of methods, such as cosine similarity, Euclidean distance, Manhattan distance, etc. In practical applications, the appropriate commonality metric method can be selected according to the specific situation. For example, if you are more concerned about the directional similarity between vectors, you can choose cosine similarity; if you are more concerned about the absolute distance between vectors, you can choose Euclidean distance.

[0179] The second commonality metric score is used to quantify the similarity between the coded representation of the true monitoring data and the coded representation of the normal monitoring data. Since the normal data represents the monitoring situation of the reservoir irrigation area under normal operation, it is essentially different from the true data (warning state), so in theory, the coded representations of the two should have a lower similarity. This score can also be calculated by methods such as cosine similarity and Euclidean distance.

[0180] Based on the first commonality metric score and the second commonality metric score, the computer system can further calculate a first error value, which reflects the performance difference of the representation coding network component when processing the true data, the noisy true data, and the normal data.

[0181] One method is to directly use the difference or ratio of the commonality metric scores as the error value. For example, if cosine similarity is used as the commonality metric method, the difference between the first commonality metric score and the second commonality metric score can be calculated. The larger the difference, the better the performance of the characterization coding network component in distinguishing between the warning state and the normal state. In order to evaluate the performance more comprehensively, it is also possible to consider weighted averaging or other combinations of the two commonality metric scores to calculate the error value.

[0182] Through the implementation of step S2231, the computer system can quantitatively characterize the performance differences of the encoding network components when processing different types of data, providing an important basis for subsequent parameter adjustment. This error calculation method based on commonality metric scoring helps to improve the robustness and generalization ability of the model, enabling it to better adapt to complex and changing monitoring environments.

[0183] As an implementation mode, step S224, characterizing and encoding the early warning reservoir irrigation area monitoring data and the noisy early warning reservoir irrigation area monitoring data respectively according to the characterization coding network component, and obtaining the early warning monitoring data coding representation corresponding to the early warning reservoir irrigation area monitoring data and the noisy early warning monitoring data coding representation corresponding to the noisy early warning reservoir irrigation area monitoring data, including:

[0184] Step S2241: characterize and encode the early warning reservoir irrigation area monitoring data according to the first characterization and encoding branch component to obtain a code representation of the early warning monitoring data;

[0185] Step S2242: Characterize and encode the noise-added warning reservoir irrigation area monitoring data according to the second characterization and encoding branch component to obtain a coded representation of the noise-added warning monitoring data.

[0186] In step S2241, the computer system uses the first characterization coding branch component to characterize and code the early warning reservoir irrigation area monitoring data to generate an early warning monitoring data code representation. The early warning reservoir irrigation area monitoring data refers to the monitoring data of the reservoir irrigation area that has triggered or is close to triggering the early warning condition, which reflects the actual situation of the reservoir irrigation area under the early warning state.

[0187] The first representation encoding branch component is a pre-trained neural network branch specifically designed to process monitoring data in the early warning state. It may contain multiple layers, such as input layer, hidden layer, and output layer, each of which contains a certain number of neurons and connection weights. These weights are gradually adjusted by the learning algorithm during the model training process to minimize the prediction error.

[0188] When the warning reservoir irrigation area monitoring data is input into the first representation coding branch component, it will undergo a series of nonlinear transformations and linear transformations and eventually be converted into a low-dimensional coding vector. This coding vector captures the key features and information under the warning state and is the basis for subsequent Hongxun predictions.

[0189] For example, assume that the early warning reservoir irrigation area monitoring data contains multiple features, such as water level, rainfall, soil moisture, etc. These data are input into the first representation encoding branch component in the form of vectors, for example: [3.8, 45, 80] (representing water level, rainfall, and soil moisture, respectively). After being processed by this branch component, the output may be a low-dimensional encoding vector, such as: [0.7, -0.4, 0.2]. This encoding vector represents the data characteristics in the early warning state and is the basis for subsequent processing and analysis.

[0190] In step S2242, the computer system uses the second characterization encoding branch component to characterize and encode the noisy early warning reservoir irrigation area monitoring data to generate a code representation of the noisy early warning monitoring data. The noisy early warning reservoir irrigation area monitoring data is the result of noise processing on the early warning reservoir irrigation area monitoring data, which aims to increase the diversity of the data and the robustness of the model by introducing noise.

[0191] The second representation encoding branch component is similar to the first representation encoding branch component. The second representation encoding branch component is also a pre-trained neural network branch, but it is specifically used to process the early warning monitoring data after adding noise. Although it has the same structure as the first branch (i.e., the same number of layers and neurons), their weights and biases may be different, which enables the two branches to produce different encoding representations when processing the same data.

[0192] Noise processing is a data enhancement technology that generates new data samples by adding random noise to the original data. In step S2242, the noise-added early warning reservoir irrigation area monitoring data is obtained by adding Gaussian noise or other types of noise to the early warning reservoir irrigation area monitoring data. This processing is intended to simulate the uncertainty and variability in the real world and improve the generalization ability of the model.

[0193] Similar to step S2241, after the noise-added warning reservoir irrigation area monitoring data is input into the second representation coding branch component, it will undergo a series of nonlinear transformations and linear transformations and finally be converted into a low-dimensional coding vector. This coding vector captures the key features and information under the noise-added warning state, but due to the influence of noise, it may be different from the coding representation of the warning monitoring data.

[0194] For example, suppose the noisy warning reservoir irrigation area monitoring data is obtained by adding Gaussian noise to the warning reservoir irrigation area monitoring data. For example, the original data is [3.8, 45, 80], and after adding noise, it may become [3.75, 44.5,79.8]. After this noisy data is input into the second representation encoding branch component, the output may be a low-dimensional encoding vector that is slightly different from the encoding representation of the warning monitoring data, such as: [0.65, -0.35, 0.25].

[0195] Through the implementation of steps S2241 and S2242, the computer system can use two independent characterization coding branch components to characterize and encode the early warning reservoir irrigation area monitoring data and the noisy early warning reservoir irrigation area monitoring data to generate corresponding coded representations. These coded representations not only capture the key features and information under the early warning state, but also increase the diversity of the data and the robustness of the model through noise processing. These coded representations will serve as the basis for subsequent Hongxun predictions, helping to improve the accuracy and reliability of the predictions. At the same time, by comparing the coded representations generated by the two branch components, the stability and consistency of the model can be further evaluated, providing a basis for the optimization and adjustment of the model.

[0196] As an implementation mode, step S2241 to step S2242, characterize and encode the early warning reservoir irrigation area monitoring data according to the first characterization coding branch component to obtain the early warning monitoring data code representation; characterize and encode the noisy early warning reservoir irrigation area monitoring data according to the second characterization coding branch component to obtain the noisy early warning monitoring data code representation, including:

[0197] Step S224a: filtering the early warning reservoir irrigation area monitoring data according to the first filter in the first characterization coding branch component to obtain a corresponding early warning reservoir irrigation area monitoring data vector set; characterizing and encoding the early warning reservoir irrigation area monitoring data vector set according to the first characterization coding unit in the first characterization coding branch component to obtain an early warning monitoring data coding representation;

[0198] Step S224b: Filter the noisy warning reservoir irrigation area monitoring data according to the second filter in the second characterization coding branch component to obtain a corresponding noisy warning reservoir irrigation area monitoring data vector set; characterize and encode the noisy warning reservoir irrigation area monitoring data vector set according to the second characterization coding unit in the second characterization coding branch component to obtain a noisy true monitoring data coding representation.

[0199] In step S224a, the computer system first uses the first filter in the first representation coding branch component to filter the early warning reservoir irrigation area monitoring data to obtain a corresponding early warning reservoir irrigation area monitoring data vector set. Then, these vector sets are input into the first representation coding unit for representation coding, and finally generate the early warning monitoring data code representation.

[0200] The first filter is a key component in the first representation encoding branch component, which is responsible for preliminary feature extraction of the input data. This filter may be a convolutional layer, a recurrent layer, or other types of filters. Its parameters (such as the weight, step size, padding, etc. of the convolution kernel) are learned during the model training process to capture local features in the input data.

[0201] When the early warning reservoir irrigation area monitoring data is input into the first filter, it slides along a specific dimension (such as the time axis, the space axis), performs weighted summation or other types of transformations on the data in the local area, and thus generates a series of feature vectors. These feature vectors capture the local patterns or structures in the data and provide a basis for subsequent representation encoding.

[0202] For example, suppose that the early warning reservoir irrigation area monitoring data is a set of time series data, containing information such as water level and rainfall at multiple monitoring points. The first filter may be a one-dimensional convolution layer with a convolution kernel size of 3, a step size of 1, and a padding of 0. When this convolution layer is applied to the early warning data, it slides along the time axis and performs a weighted sum of the data in each local window to generate a series of feature vectors. For example, for the input data [x_1, x_2, x_3, x_4, x_5], the convolution operation may produce the output [y_1, y_2, y_3, y_4], where each y_i is the weighted sum of the data in the corresponding local window.

[0203] After obtaining the early warning reservoir irrigation area monitoring data vector set, the computer system inputs it into the first representation coding unit for further representation coding. The first representation coding unit may include multiple fully connected layers, activation functions, pooling layers and other components, which work together to extract deep features in the data and generate low-dimensional encoding representations.

[0204] When the warning reservoir irrigation area monitoring data vector set is input into the first representation encoding unit, it will undergo a series of nonlinear transformations and linear transformations. These transformations may include matrix multiplication of the fully connected layer, nonlinear activation of the activation function (such as ReLU, Sigmoid, etc.), downsampling operations of the pooling layer, etc. Through these transformations, the first representation encoding unit can capture the deep features in the data and convert them into low-dimensional encoding representations.

[0205] For example, assume that the first representation encoding unit contains a fully connected layer and a ReLU activation function. When the warning reservoir irrigation area monitoring data vector set [v_1, v_2, ..., v_n] is input into this unit, the fully connected layer will perform a linear transformation on it to generate a new vector w=Wv+b, where W is the weight matrix, v is the stacked form of the vector set, and b is the bias vector. Then, the ReLU activation function applies nonlinear activation to each element in w to generate the final encoding representation e=ReLU(w). This encoding representation e is a low-dimensional vector that captures the key features of the warning reservoir irrigation area monitoring data.

[0206] In step S224b, the computer system uses a similar process as step S224a, but this time the noise-added warning reservoir irrigation area monitoring data is processed. The noise-added warning reservoir irrigation area monitoring data is the result of noise-added processing of the warning reservoir irrigation area monitoring data, which aims to increase the diversity of the data and the robustness of the model by introducing noise.

[0207] The second filter is similar to the first filter and is also a key component in the second representation encoding branch component. It is responsible for the preliminary feature extraction of the noise warning reservoir irrigation area monitoring data. Although the second filter and the first filter may be structurally the same (such as convolutional layers or recurrent layers), their parameters (such as weights, step sizes, etc.) are usually different. This difference helps the second representation encoding branch component learn a different feature representation from the first branch.

[0208] When the noisy early warning reservoir irrigation area monitoring data is input into the second filter, it undergoes a convolution or loop operation similar to the first filter to generate a series of feature vector sets. These feature vector sets capture the local features in the noisy data, and due to the influence of noise, they may be different from the feature vectors of the early warning reservoir irrigation area monitoring data.

[0209] For example, assume that the noisy early warning reservoir irrigation area monitoring data is obtained by adding Gaussian noise to the early warning data. The second filter may be a convolutional layer with the same structure as the first filter but different parameters. When this convolutional layer is applied to the noisy data, it generates a series of new feature vector sets that contain both the characteristics of the early warning data and the impact of noise. After obtaining the noisy early warning reservoir irrigation area monitoring data vector set, the computer system inputs it into the second representation coding unit for further representation coding. The second representation coding unit may be similar to the first representation coding unit in structure, but their parameters are usually different. This difference enables the two branch components to generate different encoding representations, which helps to improve the diversity and robustness of the model.

[0210] When the noisy warning reservoir irrigation area monitoring data vector set is input into the second representation coding unit, it undergoes a series of nonlinear transformations and linear transformations. These transformations are similar to those in the first representation coding unit, but due to different parameters, they generate different encoding representations. This encoding representation captures the key features of the noisy warning reservoir irrigation area monitoring data and reflects the impact of noise on data representation.

[0211] For example, suppose that the second representation encoding unit also contains a fully connected layer and a ReLU activation function, but its weight matrix and bias vector are different from those of the first representation encoding unit. When the noisy warning reservoir irrigation area monitoring data vector set is input into this unit, it generates a new encoding representation e'=ReLU(W'v'+b'), where W' and b' are the weight matrix and bias vector of the second representation encoding unit, respectively, and v' is the stacked form of the feature vector set of the noisy data. This encoding representation e' is different from the early warning monitoring data encoding representation e, and it reflects the impact of noise on data representation.

[0212] As an implementation mode, step S225, performing calculation according to the early warning monitoring data coding representation, the noisy early warning monitoring data coding representation and the normal monitoring data coding representation to obtain a second error value includes:

[0213] Step S2251: Determine a second error value based on a third commonality metric score between the early warning monitoring data coding representation and the noisy early warning monitoring data coding representation, and a fourth commonality metric score between the authentic monitoring data coding representation and the normal monitoring data coding representation.

[0214] In step S2251, the computer system calculates the third commonality metric score between the early warning monitoring data encoding representation and the noisy early warning monitoring data encoding representation, and then calculates the fourth commonality metric score between the true monitoring data encoding representation and the normal monitoring data encoding representation. These two scores respectively reflect the similarity or distance between the early warning state and the noisy early warning state, and the true early warning state and the normal state in the coding space. Based on these two scores, the computer system can further calculate a second error value for evaluating the performance difference of the characterizing coding network component when processing data of different states.

[0215] The third commonality metric score is used to quantify the similarity between the coded representation of the early warning monitoring data and the coded representation of the noisy early warning monitoring data. Since the noisy early warning data is obtained by slightly perturbing the early warning data, in theory, there should be a high similarity between the coded representation of the early warning monitoring data and the coded representation of the noisy early warning monitoring data. This similarity can be calculated using a variety of commonality metrics, such as cosine similarity, Euclidean distance, Manhattan distance, etc.

[0216] The fourth commonality metric score is used to quantify the similarity between the coded representation of the true monitoring data and the coded representation of the normal monitoring data. Since the true warning data represents the actual situation of the reservoir irrigation area under the warning state, and the normal data represents the situation of the reservoir irrigation area under normal operation, the two are essentially different. Therefore, in theory, there should be a low similarity between the coded representation of the true monitoring data and the coded representation of the normal monitoring data. This score can also be calculated by methods such as cosine similarity and Euclidean distance.

[0217] Based on the third commonality metric score and the fourth commonality metric score, the computer system can further calculate a second error value, which reflects the performance difference of the characterizing coding network component when processing the warning state, the noisy warning state, and the normal state data.

[0218] One method is to directly use the difference or ratio of the commonality metric scores as the error value. The larger the error value, the better the performance of the characterization encoding network component in distinguishing between the warning state and the normal state. In order to evaluate the performance more comprehensively, it is also possible to consider calculating the error value by weighted averaging or other combinations of the two commonality metric scores.

[0219] Through the implementation of step S2251, the computer system can quantitatively characterize the performance differences of the encoding network components when processing different state data, providing an important basis for subsequent parameter adjustment. This error calculation method based on commonality metric scoring helps to improve the robustness and generalization ability of the model, enabling it to better adapt to complex and changing monitoring environments.

[0220] As an implementation mode, the pre-debugging characterization coding network component includes a first pre-debugging characterization coding branch component and a second pre-debugging characterization coding branch component, the first pre-debugging characterization coding branch component is obtained by the first characterization coding branch component based on the label-free pre-debugging, and the second pre-debugging characterization coding branch component is obtained by the second characterization coding branch component based on the label-free pre-debugging; based on this, step S30, through the pre-debugging characterization coding network component and the Hongxun reasoning component, performs Hongxun prediction on the reservoir irrigation area monitoring data template to obtain the predicted Hongxun trigger score, including:

[0221] Step S31: The reservoir irrigation area monitoring data template is loaded into the first pre-debugging characterization coding branch component and the second pre-debugging characterization coding branch component for characterization coding, respectively, to obtain a first template monitoring data coding representation, a second template monitoring data coding representation and a third template monitoring data coding representation; the first template monitoring data coding representation is a coding array obtained by executing the first pre-debugging characterization coding branch component and performing weight adjustment at the same time, the second template monitoring data coding representation is a coding array obtained by executing the second pre-debugging characterization coding branch component and performing weight adjustment at the same time, and the third template monitoring data coding representation is a coding array obtained by executing the second pre-debugging characterization coding branch component and not performing weight adjustment at the same time;

[0222] Step S32: predicting the Hongxun trigger score of the third template monitoring data encoding representation according to the Softmax classifier in the Hongxun reasoning component, and obtaining the confidence of the third template monitoring data encoding representation on each Hongxun trigger score;

[0223] Step S33: Determine the predicted flood signal trigger score based on the confidence level.

[0224] In step S31, the computer system first loads the reservoir irrigation area monitoring data template into the first pre-debugging characterization encoding branch component and the second pre-debugging characterization encoding branch component for characterization encoding. These two branch components are obtained based on unlabeled pre-debugging, and they can extract key features and information from the monitoring data and convert them into low-dimensional encoding representations.

[0225] The first pre-debugging characterization encoding branch component is an optimized version of the first characterization encoding branch component obtained after unlabeled pre-debugging. It is specifically used to process reservoir irrigation area monitoring data and can generate a first template monitoring data encoding representation. This encoding representation not only captures the key features in the data, but also is further optimized through weight adjustment to improve the accuracy and robustness of the prediction. The second pre-debugging characterization encoding branch component is similar to the first branch component, and the second pre-debugging characterization encoding branch component is also an optimized version obtained based on unlabeled pre-debugging. It is used to generate the second template monitoring data encoding representation, and weight adjustment is also performed to optimize performance. In addition to the first and second encoding representations, step S31 also generates a coding array without weight adjustment, that is, the third template monitoring data encoding representation. This encoding representation is obtained by executing the second pre-debugging characterization encoding branch component, but it does not perform additional weight adjustment. The purpose of this step is to compare the impact of weight adjustment on the prediction results, so as to verify the effectiveness of weight adjustment.

[0226] When the reservoir irrigation area monitoring data template is input into these two branch components, they will filter, nonlinearly transform and linearly transform the data respectively, and finally generate the corresponding coded representation. These coded representations are the basis for subsequent Hongxun predictions. For example, suppose that the reservoir irrigation area monitoring data template contains information such as water level, rainfall, soil moisture, etc. at multiple monitoring points. When these data are input into the first pre-debugging representation encoding branch component, it may generate an encoding vector of the form [0.6, -0.3, 0.2] as the first template monitoring data encoding representation. Similarly, the second pre-debugging representation encoding branch component may generate another encoding vector of the form [0.5, -0.2, 0.1] as the second template monitoring data encoding representation. The third template monitoring data encoding representation without weight adjustment may be slightly different, such as [0.55, -0.25, 0.15].

[0227] In step S32, the computer system uses the Softmax classifier in the Hongxun reasoning component to predict the Hongxun trigger score for the third template monitoring data encoding representation. The Softmax classifier is a commonly used multi-classification algorithm that can map the input feature vector to different categories and give the confidence of each category.

[0228] In the Hongxun reasoning component, the Softmax classifier receives the third template monitoring data encoding representation as input, and outputs a probability distribution vector through a series of calculations and transformations. Each element in this vector represents the confidence of a Hongxun trigger score, that is, the probability that the data belongs to this score.

[0229] When the third template monitoring data encoding representation is input to the Softmax classifier, it will first pass through a series of fully connected layers for linear transformation and nonlinear activation, and then apply the Softmax function to convert the output into a probability distribution. The Hongxun trigger score corresponding to the maximum value in this probability distribution vector is the prediction result.

[0230] For example, suppose the Softmax classifier outputs a probability distribution vector of the form [0.1, 0.3, 0.6], which means that the probability that the input data belongs to the first Hongxun trigger score is 0.1, the probability that it belongs to the second score is 0.3, and the probability that it belongs to the third score is 0.6. Therefore, the predicted Hongxun trigger score is the third score.

[0231] In step S33, the computer system determines the final predicted Hongxun trigger score based on the confidence of the Softmax classifier output. This process usually involves selecting the score with the highest confidence as the prediction result.

[0232] In the probability distribution vector output by the Softmax classifier, the score with the highest confidence is the predicted Hongxun trigger score. This score reflects the highest probability that the input data belongs to this category, so it can be considered the most likely prediction result. For example, continuing the above example, the probability distribution vector output by the Softmax classifier is [0.1, 0.3, 0.6], where the third score has the highest confidence (0.6). Therefore, the computer system determines the third score as the final predicted Hongxun trigger score.

[0233] Through the implementation of step S30, the computer system can use the pre-debugging representation coding network component and the flood information reasoning component to perform flood information prediction on the reservoir irrigation area monitoring data template and obtain the predicted flood information trigger score. This score reflects the risk level of flood events in the reservoir irrigation area and provides an important basis for subsequent flood control decisions. At the same time, by comparing the first, second and third template monitoring data coding representations and their impact on the prediction results, the model performance can be further verified and optimized.

[0234] As an implementation manner, the prediction error includes a first prediction error and a second prediction error; step S40, determining the prediction error according to the predicted flood signal trigger score and the prior flood signal trigger score, includes:

[0235] Step S41: obtaining a first prediction error according to a relative entropy function value between a first template monitoring data encoding representation and a second template monitoring data encoding representation;

[0236] Step S42: Obtaining a second prediction error according to a cross entropy function value between the predicted flood signal trigger score and the prior flood signal trigger score;

[0237] Step S43: Obtain a prediction error based on the first prediction error and the second prediction error.

[0238] In step S41, the computer system calculates a first prediction error based on a relative entropy function value between the first template monitoring data encoding representation and the second template monitoring data encoding representation. The first template monitoring data encoding representation and the second template monitoring data encoding representation are generated by the first pre-debugging characterization encoding branch component and the second pre-debugging characterization encoding branch component, respectively, and they capture key features in the reservoir irrigation area monitoring data.

[0239] Relative entropy (Kullback-Leibler Divergence, KL divergence): Relative entropy is a way to measure the difference between two probability distributions. It represents the amount of information loss required when using one distribution to approximate another distribution.

[0240] When calculating the first prediction error, the computer system regards the first template monitoring data encoding representation and the second template monitoring data encoding representation as two probability distributions (the encoding representation can be converted into a probability distribution through the Softmax function), and then calculates the relative entropy between them. This value reflects the degree of difference between the two encoding representations. The greater the difference, the higher the relative entropy value.

[0241] In step S42, the computer system calculates a second prediction error based on the cross entropy function value between the predicted flood signal trigger score and the prior flood signal trigger score. The predicted flood signal trigger score is the prediction result of the flood signal reasoning component on the monitoring data, while the prior flood signal trigger score is the actual value obtained based on historical data or expert knowledge.

[0242] Cross-Entropy is another way to measure the difference between two probability distributions. It is often used in the loss function of classification problems to measure the difference between the probability distribution predicted by the model and the probability distribution of the true label. When calculating the second prediction error, the computer system treats the predicted Hongxun trigger score as the probability distribution predicted by the model, and the prior Hongxun trigger score as the probability distribution of the true label (usually a one-hot encoded vector), and then calculates the cross entropy between them. This value reflects the degree of difference between the model prediction result and the actual value. The greater the difference, the higher the cross entropy value.

[0243] In step S43, the computer system calculates the total prediction error based on the first prediction error and the second prediction error. The total prediction error is an important indicator for comprehensively measuring the performance of the model, which reflects the overall performance of the model in capturing data features and predicting actual values. The total prediction error can be obtained by weighted averaging the first prediction error and the second prediction error or other combinations thereof.

[0244] Through the implementation of step S40, the computer system can comprehensively measure the performance of the model in predicting the Hongxun trigger score. The first prediction error focuses on the differences between different branch components within the model, which helps to evaluate the diversity and robustness of the model; the second prediction error focuses on the difference between the model prediction results and the actual values, which helps to evaluate the accuracy and reliability of the model. The total prediction error is a comprehensive reflection of the two, providing an important basis for subsequent model adjustment and optimization.

[0245] As an implementation method, the reservoir irrigation area monitoring data includes a collection of reservoir tributary monitoring data; in step S200, the reservoir irrigation area monitoring data is loaded into the pre-debugging representation coding network component in the Hongxun prediction neural network for representation coding to obtain the monitoring data coding representation, including:

[0246] Step S210: Acquire the reservoir tributary monitoring data in the reservoir tributary monitoring data collection;

[0247] Step S220: Load the reservoir tributary monitoring data into the pre-debugging characterization coding network component for characterization coding to obtain a coded representation of the reservoir tributary monitoring data.

[0248] In step S210, the computer system obtains the monitoring data of the reservoir tributaries in the reservoir tributary monitoring data collection. The reservoir tributary monitoring data collection is a collection of monitoring data of multiple reservoir tributaries, each of which has its unique geographical location, hydrological characteristics and monitoring indicators. These data are usually collected in real time through a sensor network, including information on water level, rainfall, flow, water quality and other aspects.

[0249] There are many ways to collect monitoring data for reservoir tributaries, including but not limited to sensor monitoring, remote sensing monitoring, and manual observation. Sensor monitoring is one of the most commonly used methods. It collects data in real time by deploying sensors such as water level gauges and rain gauges at key locations in reservoir tributaries. Remote sensing monitoring uses remote sensing technologies such as satellites or drones to obtain hydrological information over a large area. Manual observation is when professionals go to the site to conduct field measurements at specific time points.

[0250] The collected monitoring data of reservoir tributaries usually exist in the form of time series, each of which contains a series of timestamps and corresponding monitoring values. For example, a water level time series may contain all water level records from a certain time point to the current time, and each record includes a timestamp and a water level value. In addition to the time series data, it may also contain some additional metadata, such as the location, type, calibration information, etc. of the sensor.

[0251] After obtaining the raw data, the computer system usually performs a series of preprocessing operations on the data to ensure the quality and consistency of the data. These preprocessing operations may include data cleaning (removing outliers, filling missing values, etc.), data conversion (such as unit conversion, data type conversion, etc.) and data normalization (scaling the data to the same dimension).

[0252] In step S220, the computer system loads the pre-processed reservoir tributary monitoring data into the pre-debugging representation encoding network component for representation encoding to generate a reservoir tributary monitoring data encoding representation. The pre-debugging representation encoding network component is a neural network model optimized by unlabeled pre-debugging (i.e., unsupervised learning), which can automatically extract useful features and information from the original monitoring data and convert it into a low-dimensional, efficient encoding representation.

[0253] Pre-debugging representation encoding network components consist of one or more neural network layers, such as convolutional layers, recurrent layers, fully connected layers, etc. Each layer has its own specific functions and parameters, such as convolutional layers are used to extract local features, recurrent layers are used to process time dependencies in time series data, and fully connected layers are used to map features to the encoding space. The parameters of these layers are gradually adjusted through optimization algorithms (such as gradient descent) during unsupervised learning.

[0254] When the reservoir tributary monitoring data is input into the pre-debugging representation encoding network component, it is processed through each neural network layer in turn. In each layer, the data undergoes a series of transformations and calculations, such as convolution operations, pooling operations, activation function applications, etc. These transformations are designed to extract key features and information from the data and convert them into higher-level representations. Finally, after the last fully connected layer, the data is mapped into a low-dimensional encoding space, generating an encoding vector as output.

[0255] The resulting encoded representation is a low-dimensional vector that captures key features and information in the reservoir tributary monitoring data. These features may include the changing trend of water levels, the intensity of rainfall, the fluctuation of flow, etc. The low dimensionality of the encoded representation makes it easier to be used in subsequent processing and analysis, while also reducing computational complexity and storage requirements.

[0256] For example, continuing the above example, suppose there is a pre-debugging representation encoding network component, which consists of a convolutional layer, a maximum pooling layer, and a fully connected layer. When the water level time series of Stream 1 is input into this component, it will be sequentially extracted through the convolutional layer, reduced in dimension through the maximum pooling layer, and finally mapped to the encoding space through the fully connected layer.

[0257] Through the implementation of steps S210 and S220, the computer system can successfully load the reservoir tributary monitoring data into the pre-debugging representation coding network component for representation coding, generating low-dimensional and efficient coding representations. These coding representations not only retain the key features and information in the original data, but also reduce the dimension and complexity of the data, providing strong support for subsequent Hongxun predictions.

[0258] As an implementation method, step S300, performing Hongxun prediction on the encoded representation of the monitoring data according to the Hongxun reasoning component in the Hongxun prediction neural network, and obtaining the Hongxun trigger score corresponding to the reservoir irrigation area monitoring data, includes:

[0259] Step S310: predicting the Hongxun trigger score of the reservoir tributary monitoring data encoding representation according to the Softmax classifier in the Hongxun reasoning component, and obtaining the confidence of the reservoir tributary monitoring data encoding representation on each Hongxun trigger score;

[0260] Step S320: determining the flood signal triggering score of the reservoir tributary monitoring data corresponding to each reservoir tributary monitoring data based on the confidence level;

[0261] Step S330: The flood signal triggering scores of the reservoir tributary monitoring data are integrated to obtain the flood signal triggering scores corresponding to the collection of reservoir tributary monitoring data.

[0262] In step S310, the computer system uses the Softmax classifier in the flood information reasoning component to predict the flood information trigger score for the encoded representation of the reservoir tributary monitoring data. The Softmax classifier is a commonly used multi-classification algorithm that can map the input feature vector to different categories and give the confidence of each category. In this scenario, the flood information trigger score can be regarded as different categories, and each score represents a flood risk level.

[0263] The Softmax classifier receives an input vector (i.e., the encoded representation of the reservoir tributary monitoring data) and performs linear transformation and nonlinear activation through a series of fully connected layers. In the last layer, it uses a Softmax function to convert the output into a probability distribution, i.e., the confidence of each Hongxun trigger score. The calculation formula of the Softmax function is: ; where x is the input vector, z jis the original output of the model for the jth category (i.e., the linear transformation result without the Softmax function), and K is the total number of categories. The Softmax function converts z j Convert to probability , which represents the probability that the input x belongs to the jth category.

[0264] For example, suppose there is a set of coded representations of monitoring data for three reservoir tributaries, each of which is a four-dimensional vector. After processing these coded representations, the Softmax classifier may output three probability distribution vectors, corresponding to three flood information trigger scores (for example, low, medium, and high risk). A possible output is: , , These probability distribution vectors represent the confidence that each encoded representation belongs to a different Hongxun trigger score.

[0265] In step S320, the computer system determines the flood signal trigger score of each reservoir tributary monitoring data corresponding to each reservoir tributary monitoring data based on the confidence output by the Softmax classifier. This process usually involves selecting the category with the highest confidence as the prediction result.

[0266] For each coded representation of reservoir tributary monitoring data, the computer system will select the category with the highest confidence in the probability distribution vector as the Hongxun trigger score of the data. This means that if the probability distribution vector corresponding to a certain coded representation is [0.1, 0.7, 0.2], the system will consider the Hongxun trigger score corresponding to the data to be "medium risk" (assuming that "medium risk" is the second category). For example, continuing the above example, for coded representation 1, the category corresponding to the highest confidence is "medium risk" (0.7), so its Hongxun trigger score is "medium risk". Similarly, the Hongxun trigger scores of coded representation 2 and coded representation 3 are also determined to be "medium risk" and "low risk", respectively.

[0267] In step S330, the computer system fuses the flood signal trigger scores of the monitoring data of each reservoir tributary to obtain the flood signal trigger score corresponding to the entire collection of reservoir tributary monitoring data. This process aims to comprehensively consider the data of all tributaries to obtain a more comprehensive and accurate flood risk assessment. The fusion method can be selected according to actual conditions. Common methods include simple average, weighted average, median, etc. The simple average is to add the flood signal trigger scores of all tributaries and divide it by the number of tributaries; the weighted average is to give different weights to each tributary according to its importance or reliability and then average them; the median is to sort the flood signal trigger scores of all tributaries and take the middle value.

[0268] For example, if we continue to use the above example and select simple average as the fusion method, then the flood trigger score corresponding to the entire reservoir tributary monitoring data collection will be the average of the scores of the three tributaries. The "medium / low risk" here is a simplified representation, and it may need to be quantified into a specific score value or risk level in actual applications.

[0269] Through the implementation of step S300, the computer system can use the flood prediction neural network to conduct a comprehensive flood risk assessment on the monitoring data of the reservoir tributaries. This process not only takes into account the data characteristics of a single tributary, but also integrates the data of multiple tributaries to obtain more accurate and reliable risk assessment results. This is of great significance for guiding flood prevention work and protecting people's lives and property.

[0270] As an implementation mode, step S330, the flood signal triggering scores of the reservoir tributary monitoring data are integrated to obtain the flood signal triggering scores corresponding to the collection of the reservoir tributary monitoring data, including:

[0271] Step S331: clustering the reservoir tributary monitoring data to be subjected to flood warning prediction, and determining the reservoir tributary monitoring data flood warning trigger score corresponding to the reservoir tributary monitoring data in each cluster;

[0272] Step S332: Perform preliminary average adjustment on the flood signal triggering scores of the reservoir tributary monitoring data in each cluster to obtain the flood signal triggering scores of the reservoir tributary monitoring data group corresponding to each cluster;

[0273] Step S333: perform secondary average adjustment on the flood signal triggering scores of the reservoir tributary monitoring data groups corresponding to each cluster to obtain the flood signal triggering scores corresponding to the reservoir tributary monitoring data collection.

[0274] In step S331, the computer system clusters the reservoir tributary monitoring data to be subjected to flood forecasting. The purpose of clustering is to classify the reservoir tributary monitoring data with similar characteristics or behaviors into one category so as to facilitate a more accurate risk assessment later. There are many clustering methods, including but not limited to K-means clustering, hierarchical clustering, DBSCAN, etc.

[0275] A commonly used clustering method that attempts to find K clusters such that the data points within each cluster are as similar as possible, while the data points between different clusters are as different as possible. The basic steps of K-means clustering include randomly selecting K initial centroids, assigning each data point to the nearest centroid, updating the centroid position, and repeating the assignment and update process until convergence.

[0276] For example, suppose there are 10 monitoring data points for reservoir tributaries, each of which contains multiple features such as water level, rainfall, soil moisture, etc. They can be divided into 3 clusters using the K-means clustering method. First, 3 initial centroids are randomly selected; then, each data point is assigned to the nearest cluster based on its distance from the centroid; then, the centroid position of each cluster is updated to the mean of all data points in the cluster; the assignment and update process is repeated until the centroid position no longer changes significantly. Finally, 3 clusters are obtained, and the data points in each cluster have high similarity in features.

[0277] By clustering, reservoir tributaries with similar flood risk characteristics can be grouped together, so that the flood risk within each cluster can be assessed more accurately. At the same time, clustering can also help reduce the impact of data noise and outliers and improve the robustness of risk assessment.

[0278] In step S332, the computer system performs preliminary average adjustment on the flood signal triggering score of the reservoir tributary monitoring data in each cluster. The purpose of the preliminary average adjustment is to obtain a flood signal triggering score that can represent the overall flood risk of the cluster by weighted summation and re-averaging.

[0279] In the weighted summation process, the flood trigger score of each reservoir tributary monitoring data is assigned a weight, which reflects the importance or reliability of the data point within the cluster. The weight can be determined based on a variety of factors such as the characteristics of the data point, historical performance, and monitoring accuracy. The calculation formula for the weighted summation is: ; where n is the number of data points in the cluster, w i is the weight of the ith data point, is the Hongxun trigger score of the ith data point. After the weighted summation, the computer system averages the weighted sum to obtain a single Hongxun trigger score to represent the risk level of the entire cluster. The average calculation formula is: ; For example, continuing with the above example, assume that the 10 reservoir tributary monitoring data points have been divided into 3 clusters. For the data points in each cluster, different weights can be assigned according to their historical performance and monitoring accuracy. Then, the weighted summation and averaging method is used to calculate the flood trigger score of each cluster. For example, for the first cluster, assume that it contains 3 data points, and their flood trigger scores are 0.6, 0.7 and 0.8, respectively, and the weights are 0.8, 0.9 and 0.7, respectively (the total weight is 2.4). Then the weighted sum is , and the averaged intra-cluster score is .

[0280] In step S333, the computer system performs a secondary average adjustment on the flood signal trigger scores corresponding to each cluster to obtain the flood signal trigger score corresponding to the entire collection of reservoir tributary monitoring data. The purpose of the secondary average adjustment is to comprehensively consider the risk levels of all clusters and obtain a single score that can represent the flood risk of the entire reservoir irrigation area. Similar to the preliminary average adjustment, the secondary average adjustment also obtains the final score by calculating the weighted and re-averaging the scores of all clusters. However, the difference is that in the secondary average adjustment, the weight of each cluster may no longer be based on the characteristics or performance of the data point, but on the importance or scope of influence of the cluster. The calculation formula for the secondary average is: ; where m is the number of clusters, W j is the weight of the jth cluster, is the Hongxun trigger score of the jth cluster.

[0281] For example, continuing with the above example, suppose the Hongxun trigger scores of the three clusters have been calculated to be 0.82, 0.75, and 0.68, and different weights have been assigned to them according to the importance or influence of the clusters, which are 1.2, 1.0, and 0.8 (the total weight is 2.0). Then the combined score after the second average is .

[0282] Through the implementation of step S330, the computer system can comprehensively consider the flood risks of multiple reservoir tributary monitoring data and obtain a single score that can represent the flood risk of the entire reservoir irrigation area. This score not only considers the risk level of a single tributary, but also integrates the risk characteristics of multiple tributaries through clustering and average adjustment, thereby improving the accuracy and reliability of risk assessment.

[0283] It should be noted that in the embodiments of the present application, if the above-mentioned reservoir irrigation area flood control early warning method based on big data analysis is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present application can be essentially or partly reflected in the form of a software product that contributes to the relevant technology. The software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a U disk, a mobile hard disk, a read-only memory (ROM), a disk or an optical disk. In this way, the embodiments of the present application are not limited to any specific hardware, software or firmware, or any combination of hardware, software, and firmware.

[0284] The above is only an implementation method of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A flood prevention and early warning method for reservoir irrigation areas based on big data analysis, characterized in that: The method comprises: Obtain monitoring data of reservoir irrigation areas for flood forecasting; The reservoir irrigation area monitoring data is loaded into the pre-debugging characterization coding network component in the Hongxun prediction neural network for characterization coding to obtain a coded representation of the monitoring data; the pre-debugging characterization coding network component is obtained by pre-debugging the characterization coding network component in the Hongxun prediction neural network without labels based on the pre-debugging reservoir irrigation area monitoring database; Performing Hongxun prediction on the encoded representation of the monitoring data according to the Hongxun reasoning component in the Hongxun prediction neural network to obtain a Hongxun trigger score corresponding to the monitoring data of the reservoir irrigation area; The adjustment process of the Hongxun prediction neural network includes: Obtaining a pre-debugging reservoir irrigation area monitoring database, a reservoir irrigation area monitoring data template, and a priori flood signal trigger score corresponding to the reservoir irrigation area monitoring data template; Acquire the real reservoir irrigation area monitoring data, the early warning reservoir irrigation area monitoring data and the normal reservoir irrigation area monitoring data in the pre-debugging reservoir irrigation area monitoring database; the real reservoir irrigation area monitoring data represents the reservoir irrigation area monitoring data corresponding to the early warning state, the early warning reservoir irrigation area monitoring data represents the reservoir irrigation area monitoring data obtained after the reservoir irrigation area monitoring data is transformed into the early warning state, and the normal reservoir irrigation area monitoring data represents the reservoir irrigation area monitoring data other than the real reservoir irrigation area monitoring data and the early warning reservoir irrigation area monitoring data in the pre-debugging reservoir irrigation area monitoring database; perform the label-free pre-debugging on the representation coding network component according to the real reservoir irrigation area monitoring data, the early warning reservoir irrigation area monitoring data and the normal reservoir irrigation area monitoring data to obtain the pre-debugging representation coding network component; Performing Hongxun prediction on the reservoir irrigation area monitoring data template through the pre-debugging characterization coding network component and the Hongxun reasoning component to obtain a predicted Hongxun trigger score; Determining a prediction error based on the predicted flood signal trigger score and the prior flood signal trigger score; The component parameters of the pre-debugging characterization encoding network component and / or the Hongxun reasoning component are adjusted based on the prediction error.

2. The method according to claim 1, characterized in that: The step of performing the label-free pre-debugging on the representation coding network component based on the real reservoir irrigation area monitoring data, the early warning reservoir irrigation area monitoring data and the normal reservoir irrigation area monitoring data to obtain the pre-debugging representation coding network component includes: Acquire the noisy original true reservoir irrigation area monitoring data corresponding to the original true reservoir irrigation area monitoring data and the noisy early warning reservoir irrigation area monitoring data corresponding to the early warning reservoir irrigation area monitoring data; the noisy original true reservoir irrigation area monitoring data represents the reservoir irrigation area monitoring data obtained by transforming the early warning state and / or early warning level in the original true reservoir irrigation area monitoring data but without changing the basic information; the noisy early warning reservoir irrigation area monitoring data represents the reservoir irrigation area monitoring data obtained by transforming the basic information in the early warning reservoir irrigation area monitoring data but without changing the early warning state and the early warning level; According to the representation coding network component, the true reservoir irrigation area monitoring data, the noisy true reservoir irrigation area monitoring data and the normal reservoir irrigation area monitoring data are respectively represented and coded to obtain the true monitoring data coding representation corresponding to the true reservoir irrigation area monitoring data, the noisy true monitoring data coding representation corresponding to the noisy true reservoir irrigation area monitoring data and the normal monitoring data coding representation corresponding to the normal reservoir irrigation area monitoring data; Calculating according to the true monitoring data encoding representation, the noisy true monitoring data encoding representation and the normal monitoring data encoding representation to obtain a first error value; According to the representation coding network component, the early warning reservoir irrigation area monitoring data and the noisy early warning reservoir irrigation area monitoring data are respectively represented and coded to obtain the early warning monitoring data coding representation corresponding to the early warning reservoir irrigation area monitoring data and the noisy early warning monitoring data coding representation corresponding to the noisy early warning reservoir irrigation area monitoring data; Calculating according to the early warning monitoring data coding representation, the noisy early warning monitoring data coding representation and the normal monitoring data coding representation to obtain a second error value; The component parameters of the characterization coding network component are adjusted according to the first error value and the second error value to obtain the pre-debugging characterization coding network component.

3. The method according to claim 2, characterized in that The characterization coding network component includes a first characterization coding branch component and a second characterization coding branch component, the first characterization coding branch component and the second characterization coding branch component have the same composition architecture, but have different component parameters; the characterization coding of the true reservoir irrigation area monitoring data, the noisy true reservoir irrigation area monitoring data and the normal reservoir irrigation area monitoring data are respectively performed according to the characterization coding network component to obtain the true monitoring data encoding representation corresponding to the true reservoir irrigation area monitoring data, the noisy true monitoring data encoding representation corresponding to the noisy true reservoir irrigation area monitoring data and the normal monitoring data encoding representation corresponding to the normal reservoir irrigation area monitoring data, including: Characterize and encode the real reservoir irrigation area monitoring data according to the first characterization and encoding branch component to obtain the coded representation of the real monitoring data; Characterize and encode the noisy true reservoir irrigation area monitoring data according to the second characterization and encoding branch component to obtain the coded representation of the noisy true monitoring data; The normal reservoir irrigation area monitoring data is characterized and encoded according to the first characterization coding branch component or the second characterization coding branch component to obtain a coded representation of the normal monitoring data.

4. The method according to claim 3, characterized in that The characterization and encoding of the true reservoir irrigation area monitoring data according to the first characterization and encoding branch component to obtain the coded representation of the true monitoring data; the characterization and encoding of the noisy true reservoir irrigation area monitoring data according to the second characterization and encoding branch component to obtain the coded representation of the noisy true monitoring data, includes: Filtering the real reservoir irrigation area monitoring data according to the first filter in the first representation coding branch component to obtain a corresponding real reservoir irrigation area monitoring data vector set; Performing representation encoding on the true reservoir irrigation area monitoring data vector set according to the first representation encoding unit in the first representation encoding branch component to obtain the true monitoring data encoding representation; Filtering the noisy true reservoir irrigation area monitoring data according to the second filter in the second representation coding branch component to obtain a corresponding noisy true reservoir irrigation area monitoring data vector set; The noisy true reservoir irrigation area monitoring data vector set is characterized and encoded according to the second characterization encoding unit in the second characterization encoding branch component to obtain the noisy true monitoring data encoding representation.

5. The method according to claim 2, characterized in that: The step of calculating according to the true monitoring data encoding representation, the noisy true monitoring data encoding representation and the normal monitoring data encoding representation to obtain a first error value includes: Determining the first error value based on a first commonality metric score between the authentic monitoring data coded representation and the noisy authentic monitoring data coded representation, and a second commonality metric score between the authentic monitoring data coded representation and the normal monitoring data coded representation; The characterization and encoding of the early warning reservoir irrigation area monitoring data and the noisy early warning reservoir irrigation area monitoring data are performed respectively according to the characterization and encoding network component to obtain the early warning monitoring data encoding representation corresponding to the early warning reservoir irrigation area monitoring data and the noisy early warning monitoring data encoding representation corresponding to the noisy early warning reservoir irrigation area monitoring data, including: Characterizing and encoding the early warning reservoir irrigation area monitoring data according to the first characterizing and encoding branch component to obtain a coded representation of the early warning monitoring data; The noise-added warning reservoir irrigation area monitoring data is characterized and encoded according to the second characterization and encoding branch component to obtain the encoded representation of the noise-added warning monitoring data.

6. The method according to claim 5, characterized in that The characterization and encoding of the early warning reservoir irrigation area monitoring data is performed according to the first characterization and encoding branch component to obtain the coded representation of the early warning monitoring data; Characterizing and encoding the noise-added warning reservoir irrigation area monitoring data according to the second characterization and encoding branch component to obtain the coding representation of the noise-added warning monitoring data, including: Filtering the early warning reservoir irrigation area monitoring data according to the first filter in the first characterization coding branch component to obtain a corresponding early warning reservoir irrigation area monitoring data vector set; characterizing and encoding the early warning reservoir irrigation area monitoring data vector set according to the first characterization coding unit in the first characterization coding branch component to obtain the early warning monitoring data coding representation; The noisy early warning reservoir irrigation area monitoring data is filtered according to the second filter in the second characterization coding branch component to obtain a corresponding noisy early warning reservoir irrigation area monitoring data vector set; the noisy early warning reservoir irrigation area monitoring data vector set is characterized and encoded according to the second characterization coding unit in the second characterization coding branch component to obtain the noisy true monitoring data encoding representation.

7. The method according to claim 2, characterized in that The step of calculating according to the early warning monitoring data coding representation, the noise-added early warning monitoring data coding representation and the normal monitoring data coding representation to obtain a second error value comprises: The second error value is determined based on a third commonality metric score between the early warning monitoring data coding representation and the noisy early warning monitoring data coding representation, and a fourth commonality metric score between the authentic monitoring data coding representation and the normal monitoring data coding representation.

8. The method according to any one of claims 1 to 7, characterized in that: The pre-debugging characterization coding network component includes a first pre-debugging characterization coding branch component and a second pre-debugging characterization coding branch component, wherein the first pre-debugging characterization coding branch component is obtained by the first characterization coding branch component based on the tagless pre-debugging, and the second pre-debugging characterization coding branch component is obtained by the second characterization coding branch component based on the tagless pre-debugging; The method of performing Hongxun prediction on the reservoir irrigation area monitoring data template through the pre-debugging representation coding network component and the Hongxun reasoning component to obtain a predicted Hongxun trigger score includes: The reservoir irrigation area monitoring data template is loaded into the first pre-debugging characterization coding branch component and the second pre-debugging characterization coding branch component for characterization coding, respectively, to obtain a first template monitoring data coding representation, a second template monitoring data coding representation and a third template monitoring data coding representation; the first template monitoring data coding representation is a coding array obtained by executing the first pre-debugging characterization coding branch component and performing weight adjustment at the same time, the second template monitoring data coding representation is a coding array obtained by executing the second pre-debugging characterization coding branch component and performing weight adjustment at the same time, and the third template monitoring data coding representation is a coding array obtained by executing the second pre-debugging characterization coding branch component and not performing weight adjustment at the same time; According to the Softmax classifier in the Hongxun reasoning component, the Hongxun trigger score prediction is performed on the third template monitoring data encoding representation to obtain the confidence of the third template monitoring data encoding representation on each Hongxun trigger score; The predicted flood signal trigger score is determined based on the confidence level.

9. The method according to claim 1, characterized in that: The reservoir irrigation area monitoring data includes a collection of reservoir tributary monitoring data; the pre-debugging representation coding network component of the Hongxun prediction neural network is loaded with the reservoir irrigation area monitoring data for representation coding to obtain the monitoring data coding representation, including: Acquire the reservoir tributary monitoring data in the reservoir tributary monitoring data collection; Loading the reservoir tributary monitoring data into the pre-debugging characterization coding network component for characterization coding to obtain a coded representation of the reservoir tributary monitoring data; The method of performing Hongxun prediction on the encoded representation of the monitoring data according to the Hongxun reasoning component in the Hongxun prediction neural network to obtain the Hongxun trigger score corresponding to the monitoring data of the reservoir irrigation area includes: According to the Softmax classifier in the Hongxun reasoning component, the Hongxun trigger score prediction is performed on the encoded representation of the reservoir tributary monitoring data to obtain the confidence of the encoded representation of the reservoir tributary monitoring data on each Hongxun trigger score; Determine, based on the confidence level, a flood signal trigger score for the reservoir tributary monitoring data corresponding to each of the reservoir tributary monitoring data; Clustering the reservoir tributary monitoring data to be used for flood warning prediction, and determining the reservoir tributary monitoring data flood warning trigger score corresponding to the reservoir tributary monitoring data in each cluster; Performing a preliminary average adjustment on the flood signal triggering score of the reservoir tributary monitoring data in each cluster to obtain the flood signal triggering score of the reservoir tributary monitoring data group corresponding to each cluster; The flood signal triggering score of the reservoir tributary monitoring data group corresponding to each cluster is subjected to secondary average adjustment to obtain the flood signal triggering score corresponding to the reservoir tributary monitoring data collection.

Images