A method for water resource allocation in reservoir irrigation areas based on big data analysis

Abstract

The present invention relates to the field of data processing, and specifically provides a reservoir irrigation area water resource allocation method based on big data analysis, including: obtaining allocation details information of a reservoir template and an irrigation demand training label, obtaining the water resource allocation sequence characteristics of the reservoir template and the basin basic characteristics of the basin template corresponding to the reservoir template from the allocation details information of the reservoir template; using a basic water resource allocation algorithm, according to the water resource allocation sequence characteristics and the basin basic characteristics, the irrigation demand of the irrigation area template is estimated to obtain the irrigation demand estimation information of the irrigation area template; according to the irrigation demand training label and the irrigation demand estimation information, the basic water resource allocation algorithm is trained to obtain a target water resource allocation algorithm. The present invention can improve the accuracy of the training template, thereby improving the accuracy of the algorithm in predicting irrigation demand.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method for allocating water resources in a reservoir irrigation area based on big data analysis. Background Art

[0002] In water resource management, accurate prediction and reasonable allocation of water resources are crucial to ensuring water supply for agricultural irrigation, industrial production and residents' daily life. Traditional water resource allocation methods often rely on empirical judgment and simple statistical models, which have limitations in dealing with complex and changing water resource data and are difficult to accurately reflect the dynamic relationship between water resource allocation and irrigation demand.

[0003] With the rapid development of big data analysis and machine learning technologies, more and more researchers have begun to explore the application of these advanced technologies in water resources management. However, existing water resources allocation methods based on machine learning still have some shortcomings. On the one hand, these methods usually only consider a single type of water resources data (such as historical water storage, rainfall, etc.), ignoring the multimodal characteristics of the water resources allocation process, such as time series data (such as water storage change series), spatial distribution data (such as irrigation node distribution) and static environmental data (such as watershed area, soil type, etc.). This single-modal data processing method limits the generalization ability and prediction accuracy of the model. On the other hand, existing machine learning models often require additional algorithm configuration and preprocessing of different types of feature data during training, which not only increases the complexity and training cost of the model, but also may lead to information loss and degradation of model performance. In addition, due to the heterogeneity between different feature data, how to effectively integrate these multimodal features in the training process to avoid the accumulation of training costs is also an urgent problem to be solved in the current field of water resources allocation. Summary of the invention

[0004] In view of this, an embodiment of the present invention provides a method for allocating water resources in a reservoir irrigation area based on big data analysis. The technical solution of the embodiment of the present invention is implemented as follows:

[0005] A method for allocating water resources in a reservoir irrigation area based on big data analysis comprises: obtaining allocation details information of a target reservoir when allocating water resources in a target irrigation area; obtaining water resource allocation sequence characteristics of the target reservoir and basic watershed characteristics of a target watershed corresponding to the target reservoir from the allocation details information of the target reservoir; using a target water resource allocation algorithm, in combination with the water resource allocation sequence characteristics of the target reservoir and the basic watershed characteristics of the target watershed, estimating a demand area for the target irrigation area, and obtaining demand area estimation information for the target irrigation area, wherein the demand area estimation information is an estimation result representing that the target irrigation area includes an irrigation demand area. The target water resource allocation algorithm is trained by the following steps: obtaining allocation details of the reservoir template and the irrigation demand training label of the irrigation district template corresponding to the allocation details information, wherein the irrigation demand training label is a training label representing that the irrigation district template contains an irrigation demand area; obtaining the water resource allocation sequence characteristics of the reservoir template and the basin basic characteristics of the basin template corresponding to the reservoir template from the allocation details information of the reservoir template; using the basic water resource allocation algorithm, combined with the water resource allocation sequence characteristics and the basin basic characteristics, the irrigation demand of the irrigation district template is estimated to obtain the irrigation demand estimation information of the irrigation district template; the irrigation demand estimation information is an estimation result representing that the irrigation district template contains an irrigation demand area; combining the irrigation demand training label and the irrigation demand estimation information, the basic water resource allocation algorithm is trained to obtain the target water resource allocation algorithm; the degree of fit between the irrigation demand optimization result of the irrigation district template output by the target water resource allocation algorithm and the irrigation demand training label is higher than the degree of fit between the irrigation demand estimation information and the irrigation demand training label.

[0006] Technical effect of the present invention: In the present invention, the irrigation demand training label, the water resource allocation sequence characteristics in the allocation details information of the reservoir template, and the basin basic characteristics of the basin template are used as training templates to train the basic water resource allocation algorithm. The multimodal characteristics of the water resource allocation sequence characteristics and the basin basic characteristics are used to train the basic water resource allocation algorithm. The information is more comprehensive. In the process, there is no need to perform additional configuration of the corresponding algorithm for each modal information, and the accumulation of training costs is eliminated. The trained target water resource allocation algorithm is more robust and the algorithm is more streamlined. Furthermore, the degree of agreement between the irrigation demand optimization result of the irrigation area template output by the target water resource allocation algorithm and the irrigation demand training label is higher than the degree of agreement between the irrigation demand estimation information and the irrigation demand training label. The degree of agreement between the irrigation demand optimization result and the irrigation demand training label can reflect the accuracy of the target water resource allocation algorithm in identifying the irrigation demand area, and the degree of agreement between the irrigation demand estimation information and the irrigation demand training label can reflect the accuracy of the basic water resource allocation algorithm in identifying the irrigation demand area. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 A schematic diagram of the implementation process of a reservoir irrigation area water resources allocation method based on big data analysis provided in an embodiment of the present invention.

[0008] Figure 2 A schematic diagram of the training process of the target water resource allocation algorithm provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0009] The embodiment of the present invention provides a method for allocating water resources in a reservoir irrigation area 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 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 water resources allocation method based on big data analysis provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the method includes:

[0011] Step S110: Obtaining detailed information on water resource allocation of the target reservoir in the target irrigation area.

[0012] In step S110, the computer system obtains the water resource allocation details of the target reservoir in the target irrigation area from multiple channels. These channels include, for example, reservoir management systems, irrigation area management systems, meteorological monitoring stations, hydrological monitoring stations, etc. For example, the reservoir management system will record key data such as the reservoir's water storage, outflow, and inflow; the irrigation area management system will record information such as irrigation demand, irrigation area, and irrigation cycle for each irrigation area. Meteorological monitoring stations and hydrological monitoring stations provide natural environmental data such as rainfall, evaporation, and river water levels, which are essential for understanding the natural replenishment and consumption of water resources. Secondly, regarding the content of the data, the allocation details information should cover multiple aspects to ensure the comprehensiveness and accuracy of subsequent analysis. Specifically, this information includes, but is not limited to:

[0013] Reservoir water storage: refers to the amount of water stored in a reservoir at a certain moment, which can be measured in cubic meters. For example, the water storage capacity of a reservoir at a certain moment is 100,000,000 cubic meters. Outflow: refers to the amount of water released by a reservoir to an irrigation district or other water users in a certain period of time. For example, the outflow of a reservoir on a certain day is 5,000,000 cubic meters. Inflow: refers to the amount of water received by a reservoir from natural recharge (such as rainfall, upstream water) or artificial recharge (such as water diversion projects) in a certain period of time. For example, the inflow of a reservoir on a certain day is 3,000,000 cubic meters. Irrigation demand: refers to the irrigation demand of an irrigation district in different time periods, which can be measured in cubic meters / day or cubic meters / mu. For example, the irrigation demand of an irrigation district during the peak summer season is 1,000 cubic meters / day / mu. Irrigation area: refers to the total area of ​​an irrigation district, which can be measured in mu or hectare. For example, the total area of ​​an irrigation district is 10,000 mu. Irrigation cycle: Indicates the time period of irrigation in the irrigation area, such as daily, weekly or monthly. For example, an irrigation area adopts a weekly irrigation cycle. Meteorological and hydrological data: Including rainfall, evaporation, river water level, etc., these data are very important for understanding the natural changes of water resources. For example, the rainfall on a certain day is 50 mm and the evaporation is 20 mm. In terms of data format, these distribution details can exist in the form of structured data, such as database tables, CSV files, or JSON objects. Each format has its own specific organization method, but usually contains field names and corresponding data values. For example, in a CSV file, each row represents a record, and each column represents a field, such as "date", "reservoir water storage", "discharge", etc. In a JSON object, the data exists in the form of key-value pairs, such as {"date":"2023-05-01","reservoir water storage":100000000,"discharge":5000000}. After obtaining these data, the computer system can perform data preprocessing to ensure the accuracy and consistency of the data. Preprocessing steps include data cleaning (such as removing duplicate records and handling missing values), data conversion (such as converting strings to numeric types), and data normalization (such as converting data of different dimensions to a unified dimension). For example, if the "reservoir storage capacity" field of a record is found to be empty, the system will interpolate and fill it based on the data of adjacent time points; if the data unit of the "rainfall" field is found to be millimeters, and the system's data unit is cubic meters, unit conversion is required.

[0014] Step S120: Acquire the water resource allocation sequence characteristics of the target reservoir and the basic characteristics of the target watershed corresponding to the target reservoir from the allocation details information of the target reservoir.

[0015] The water resources allocation sequence characteristics reflect the water resources allocation of the target reservoir at different time points and are time series data. They include but are not limited to:

[0016] Water storage change series: records the changes in water storage of the target reservoir at different time points, such as daily, weekly or monthly water storage. For example, a time series data point may be represented as (date, water storage), such as (2023-05-01, 100,000,000 m³).

[0017] Discharge sequence: represents the discharge of the target reservoir at different time points, that is, the amount of water released to the irrigation district or other water users. For example, a data point may be represented as (date, discharge), such as (2023-05-01, 5,000,000m³).

[0018] Inflow sequence: records the inflow of the target reservoir at different time points, including natural recharge and artificial recharge. For example, a data point may be represented as (date, inflow), such as (2023-05-01, 3,000,000m³).

[0019] The basic characteristics of the watershed describe the basic conditions of the geography, climate, hydrology, etc. of the watershed where the target reservoir is located. They are static data. They include but are not limited to:

[0020] Basin Area: Indicates the total area of ​​the basin, which can be measured in square kilometers. For example, Basin Area = 10,000km². Average Annual Precipitation: Indicates the average annual rainfall in the basin over the years, which can be measured in millimeters. For example, Average Annual Precipitation = 800mm. River Length: Indicates the total length of the main rivers in the basin, which can be measured in kilometers. For example, River Length = 500km. Soil Type: Describes the main soil types in the basin, such as sand, loam, clay, etc. This is a classification feature, which can be represented by One-Hot Encoding. Vegetation Coverage: Indicates the proportion of vegetation coverage in the basin, which can be expressed as a percentage. For example, Vegetation Coverage = 60%.

[0021] For time series data, such as water storage change series, outflow series, and inflow series, they can exist in the form of arrays or time series objects. Each data point contains a timestamp and a corresponding value. For static data, such as watershed area, average annual precipitation, etc., they can exist in the form of scalar values ​​or vectors, depending on the needs of the algorithm model.

[0022] In the process of extracting features, the computer system, for example, performs some calculations or conversions. For example, for time series data, the system calculates statistics such as sliding averages, cumulative sums, and differences to capture trends and periodicity of the data. For static data, the system performs normalization or standardization to ensure that features of different dimensions can be compared on the same scale. For example, to calculate the sliding average, the computer system can use the following formula: ; where x i is the value of the time series data at time point i, N is the size of the sliding window, and t is the current time point. By calculating the sliding average, the system can smooth the time series data and reduce noise interference. In addition, for categorical features (such as soil type), the computer system uses one-hot encoding for conversion, for example. One-hot encoding is a method of converting categorical variables into binary vectors, where each category corresponds to a position in the vector, which is 1, and the rest are 0. For example, if there are three soil types (sand, loam, clay), they can be encoded as: sand: [1,0,0]; loam: [0,1,0]; clay: [0,0,1].

[0023] Step S130: adopt the target water resource allocation algorithm, combine the water resource allocation sequence characteristics of the target reservoir and the basic characteristics of the target watershed, estimate the demand area of ​​the target irrigation area, and obtain the demand area estimation information of the target irrigation area. The demand area estimation information is an estimation result representing the irrigation demand area included in the target irrigation area.

[0024] In step S130, the target water resource allocation algorithm used by the computer system is a specially trained machine learning model that can learn the complex relationship between irrigation demand and water resource allocation from the input features. This algorithm may be based on supervised learning, which means that during the training process, it needs to use a data set with irrigation demand labels to learn how to predict irrigation demand areas based on the input features.

[0025] A feasible option is to use neural network models, such as recurrent neural networks (RNN) or long short-term memory networks (LSTM), to process water resource allocation sequence features, because they are good at processing time series data and can capture the temporal dependencies in the data. At the same time, for the basic characteristics of the watershed, fully connected layers or convolutional layers can be used to extract useful information from static data.

[0026] Before inputting the water resources allocation sequence characteristics and basin basic characteristics into the target water resources allocation algorithm, the computer system processes these characteristics appropriately to ensure that they can be processed by the model in a compatible manner.

[0027] For water resource allocation sequence features, such as water storage change sequence, outflow sequence and inflow sequence, they exist in the form of time series. Computer systems can standardize these sequences to ensure that features of different dimensions can be compared on the same scale. In addition, in order to capture the time dependency in the sequence, a sliding window technique can be used to convert the sequence data into a series of fixed-length subsequences, each of which contains data at multiple time points.

[0028] For basic features of a watershed, such as watershed area, average annual precipitation, etc., which are static and do not change over time, the computer system can convert these features into vector form, which can be used directly if the features themselves are scalar values; if the features are multidimensional (such as soil type), they can be converted into vectors using, for example, one-hot encoding or other encoding methods.

[0029] When combining the two types of features, the computer system can splice them into a large feature vector, where the time series features may exist in the form of subsequences, and the static features exist in the form of a single vector. This combined feature vector will be used as the input of the target water resource allocation algorithm. Once the feature vector is prepared, the computer system can input it into the target water resource allocation algorithm to estimate the irrigation demand area. This process can be divided into the following steps:

[0030] 1. Forward propagation: The algorithm receives the input feature vector and propagates it forward through its internal network structure (such as neural network layers). During the forward propagation process, the algorithm calculates the output of each layer until it reaches the output layer.

[0031] 2. Output layer processing: In the output layer, the algorithm will calculate the estimated value of the irrigation demand area based on the results of the forward propagation. This estimated value may be a probability distribution, indicating the probability of irrigation demand in different areas of the target irrigation area; or it may be a binary result, indicating which areas of the target irrigation area have irrigation demand and which do not.

[0032] 3. Post-processing: Depending on the specific design of the algorithm, the results of the output layer can be post-processed. For example, if the output is a probability distribution, the computer system sets a threshold to convert it into a binary result; if the output is already a binary result, smoothing or denoising is performed to improve the accuracy of the estimate.

[0033] The output of step S130 is demand area estimation information, which represents the estimated result of the target irrigation area containing the irrigation demand area. This result can be presented in the form of a spatial distribution map, where different colored areas represent different irrigation demand levels. For example, a red area may represent a high irrigation demand area, a yellow area represents a medium irrigation demand area, and a green area represents a low irrigation demand area or a no demand area.

[0034] The computer system can also calculate some statistical indicators based on the estimated information of the demand area, such as the total amount of irrigation demand, the area of ​​irrigation demand, etc., in order to provide more comprehensive information support for decision makers. These statistical indicators can help decision makers understand the overall irrigation demand of the target irrigation area and the differences in irrigation demand in different areas.

[0035] For example, assuming there is a target reservoir A and a target irrigation district B, the water resource allocation sequence characteristics of reservoir A (such as water storage change sequence, outflow sequence) and the basic characteristics of basin C (such as basin area, average annual precipitation) have been extracted through step S120. Now, it is necessary to use the target water resource allocation algorithm to estimate the irrigation demand area of ​​irrigation district B.

[0036] First, the computer system preprocesses and combines these features to generate a feature vector that contains time series features and static features. Then, this feature vector is input into the target water resource allocation algorithm for forward propagation. The algorithm may contain multiple neural network layers to extract complex relationships between features. After forward propagation, the algorithm generates a probability distribution map at the output layer, which represents the probability of irrigation demand in different areas of irrigation area B.

[0037] Finally, the computer system generates a spatial distribution map based on this probability distribution map as the estimated information of the demand area. In this distribution map, we can see that the red areas are concentrated in certain parts of irrigation area B, indicating that these areas have high irrigation demand; while the green areas are distributed in other parts, indicating that these areas have low or no irrigation demand. Decision makers can formulate targeted water resource allocation plans based on this distribution map to meet the irrigation needs of different regions.

[0038] Please refer to Figure 2 , is a schematic diagram of the training process of the target water resource allocation algorithm provided by an embodiment of the present invention, and the training is performed using the following steps:

[0039] Step S11: obtaining allocation details of the reservoir template and the irrigation demand training label of the irrigation district template corresponding to the allocation details, wherein the irrigation demand training label is a training label representing that the irrigation district template contains an irrigation demand area.

[0040] In step S11, the computer system collects data from multiple sources. These data sources include, for example, reservoir management systems, irrigation district management systems, historical meteorological records, hydrological monitoring data, etc. For example, the reservoir management system can provide detailed information on the distribution of water storage, outflow, inflow, etc. at different time points in the reservoir; the irrigation district management system can provide data such as the irrigation area, irrigation cycle, and historical irrigation records of the irrigation district; historical meteorological records and hydrological monitoring data can provide natural environment information such as rainfall, evaporation, and river flow. The collected data mainly includes:

[0041] 1. Allocation details of reservoir templates: This information records in detail the water resource allocation of the reservoir at different time points. For example, the allocation details of a reservoir template include data such as water storage, outflow, and inflow at a series of time points. These data can exist in the form of a time series, with each time point corresponding to a data point, containing the specific value of the time point. For example, a data point may be expressed as (date, water storage, outflow, inflow), such as (2023-05-01, 100,000,000m³, 5,000,000m³, 3,000,000m³).

[0042] 2. Irrigation demand training labels for irrigation district templates: These labels correspond to the allocation details of the reservoir templates and are used to indicate the presence or absence of irrigation demand or the amount of irrigation demand at different time points in the irrigation district templates. Irrigation demand training labels exist in the form of classification or regression, for example. For example, in the form of classification, the label may be a binary value, indicating whether the irrigation district template has irrigation demand at a certain point in time (such as 1 for demand and 0 for no demand); in the form of regression, the label may be a numerical value, indicating the specific irrigation demand of the irrigation district template at a certain point in time (such as 1,000,000 m³).

[0043] Step S12: From the allocation details information of the reservoir template, the water resources allocation sequence characteristics of the reservoir template and the watershed basic characteristics of the watershed template corresponding to the reservoir template are obtained.

[0044] The water resource allocation sequence feature of the reservoir template reflects the water resource allocation of the reservoir at different time points and is time series data. It includes but is not limited to the water storage change sequence, the outflow sequence, the inflow sequence, etc. For example, the water storage change sequence may be represented as a series of tuples of timestamps and corresponding water storage, such as [(2023-05-01,100,000,000m³), (2023-05-02,102,000,000m³),...]. These sequence features can reveal the changing trend and periodicity of the reservoir water resources.

[0045] The basic characteristics of the watershed describe the basic conditions of the geography, climate, hydrology, etc. of the watershed where the reservoir is located. They are static data. They include but are not limited to the watershed area, average annual precipitation, river length, soil type, vegetation coverage, etc. For example, the watershed area may be expressed as 10,000km² and the average annual precipitation may be expressed as 800mm. These basic characteristics provide the algorithm with macro-background information of the watershed, which helps to more accurately predict irrigation demand.

[0046] During the feature extraction process, the computer system performs some calculations or conversions. For example, for time series data, the system calculates statistics such as sliding averages, cumulative sums, and differences to capture trends and periodicity of the data. For static data, the system performs normalization or standardization to ensure that features of different dimensions can be compared on the same scale.

[0047] In addition, when the two types of features are used in combination, the computer system, for example, designs a feature fusion mechanism. Since time series data and static data are quite different in form and nature, direct combination can lead to information loss or misleading. Therefore, a feasible approach is to use feature encoding technology to convert static data into a form compatible with time series data. For example, categorical features (such as soil type) can be converted into binary vectors using One-Hot Encoding; or continuous features (such as watershed area) can be mapped into a low-dimensional vector space using embedding technology. For example, suppose there is a data set of allocation details of a reservoir template, which contains data such as water storage, outflow, and inflow at multiple time points. At the same time, it is also known that the reservoir is located in a specific watershed with specific geographical, climatic, and hydrological characteristics.

[0048] In step S12, the computer system first extracts the water storage change sequence from the allocation detail information as an example of the water resource allocation sequence feature: water storage change sequence = [(2023-05-01, 100,000,000m³), (2023-05-02, 102,000,000m³), ...].

[0049] Then, the basin area and average annual precipitation are extracted from the basin basic information as two examples of basin basic characteristics: basin area = 10,000km²; average annual precipitation = 800mm. In order to combine the static basin basic characteristics with the time series water resources allocation characteristics, the computer system may convert the basin area and average annual precipitation into a vector form with the same length as the time series (for example, by copying): basin area vector = [10,000km², 10,000km², 10,000km², ...]; average annual precipitation vector = [800mm, 800mm, 800mm, ...].

[0050] Finally, these features are input into the target water resource allocation algorithm for training to learn the complex relationship between irrigation demand and water resource allocation and basic characteristics of the watershed.

[0051] Step S13: adopting the basic water resource allocation algorithm, combining the water resource allocation sequence characteristics and the basic characteristics of the watershed, the irrigation demand of the irrigation area template is estimated to obtain the irrigation demand estimation information of the irrigation area template; the irrigation demand estimation information is the estimation result representing the irrigation demand area included in the irrigation area template.

[0052] In step S13, the basic water resource allocation algorithm used by the computer system can be a simple machine learning model, such as linear regression, decision tree or random forest, or a more complex neural network model. The choice of this algorithm depends on the complexity of the problem and the characteristics of the available data. Regardless of which algorithm is selected, the goal is to learn the relationship between the characteristics of the water resource allocation sequence and the basic characteristics of the basin and irrigation demand.

[0053] Taking the neural network model as an example, it may contain multiple layers, such as input layer, hidden layer and output layer. The input layer receives the water resource allocation sequence characteristics and basin basic characteristics as input; the hidden layer learns the complex relationship between the characteristics through nonlinear transformation; and the output layer outputs the estimated value of irrigation demand.

[0054] Before inputting the water resource allocation sequence features and basin basic features into the basic water resource allocation algorithm, the computer system processes these features appropriately to ensure that they can be effectively used by the algorithm. For water resource allocation sequence features, such as water storage change series, storage volume series, etc., they are, for example, time series data. The computer system, for example, converts these sequence data into a form suitable for algorithm processing, such as using sliding window technology to convert sequence data into subsequences of fixed length, or using embedding technology to map sequence data into a low-dimensional vector space. For basin basic features, such as basin area, average annual precipitation, etc., they are, for example, static data. The computer system can directly use these features as input to the algorithm, or use them in combination with other features to improve the accuracy of the estimate.

[0055] When combining two types of features, the computer system can use methods such as feature concatenation, feature fusion, or feature interaction. For example, the water resource allocation sequence features and the basin basic features can be concatenated into a long vector as the input of the algorithm; or the feature fusion technology can be used to combine the two types of features at a deeper level; the feature interaction technology can also be used to capture the nonlinear relationship between the features. Once the features are prepared, the computer system can input them into the basic water resource allocation algorithm for irrigation demand estimation. Please refer to the above description.

[0056] The output of step S13 is an estimated result representing the irrigation demand area contained in the irrigation district template. This result can exist in the form of a numerical value or a probability distribution, indicating the degree of irrigation demand in different areas of the irrigation district template. For example, in the output layer of the neural network model, a sigmoid function may be used to map the output value to the range of [0,1] to represent the probability of irrigation demand. A higher output value may indicate that the area has a higher irrigation demand, while a lower output value may indicate a lower or no irrigation demand. The computer system can also calculate some statistical indicators based on the irrigation demand estimation information, such as the total amount of irrigation demand, the area of ​​the irrigation demand area, etc., in order to provide more comprehensive information support for decision makers. These statistical indicators can help decision makers understand the overall irrigation demand situation of the irrigation district template and the differences in irrigation demand in different areas.

[0057] For example, suppose there is an irrigation district template A, which is located in a specific watershed. Through step S12, the water resource allocation sequence characteristics and the basic characteristics of the watershed have been extracted from the allocation details of the reservoir template. Now, the computer system uses a basic water resource allocation algorithm (such as a neural network model) to process these characteristics. First, it concatenates the water resource allocation sequence characteristics and the basic characteristics of the watershed into a long vector as the input of the algorithm. Then, the algorithm calculates the estimated irrigation demand through forward propagation. In this process, the algorithm can learn the complex relationship between the changes in water storage, the amount of water discharged, and the irrigation demand.

[0058] Assume that the model output layer uses a sigmoid function to map the output value to the range of [0,1]. For a certain area B in the irrigation area template A, the algorithm may output a value close to 1 (such as 0.95), indicating that the area has a high irrigation demand. For another area C, the algorithm may output a value close to 0 (such as 0.05), indicating that the irrigation demand in this area is low or no demand. Based on these estimated results, the computer system can generate an irrigation demand distribution map to intuitively display the irrigation demand of different areas in the irrigation area template A. Decision makers can formulate targeted water resource allocation plans based on this distribution map to meet the irrigation needs of different areas.

[0059] Step S14: Combine the irrigation demand training label and the irrigation demand estimation information to train the basic water resource allocation algorithm to obtain the target water resource allocation algorithm; the degree of consistency between the irrigation demand optimization result of the irrigation area template output by the target water resource allocation algorithm and the irrigation demand training label is higher than the degree of consistency between the irrigation demand estimation information and the irrigation demand training label.

[0060] In step S14, the computer system uses the irrigation demand training labels to supervise the training process of the basic water resource allocation algorithm. These training labels represent the actual irrigation demand of the irrigation area template at different time points and are the target of algorithm learning. The training process includes the following steps:

[0061] 1. Data preparation: The computer system matches the irrigation demand training labels and the irrigation demand estimation information (generated by the basic water resource allocation algorithm) to ensure that each set of training data contains input features (water resource allocation sequence features and basin basic features), estimated output (irrigation demand estimation information) and true labels (irrigation demand training labels).

[0062] 2. Loss function calculation: In order to measure the difference between the irrigation demand estimated by the algorithm and the actual irrigation demand, the computer system defines a loss function. Possible loss functions include mean square error (MSE), cross entropy loss, etc. For example, if the irrigation demand estimation information is a probability distribution, the cross entropy loss can be used to measure the difference between the estimated probability distribution and the actual label.

[0063] 3. Back propagation: After calculating the loss function, the computer system uses the back propagation algorithm to transfer the loss value from the output layer back to the input layer layer by layer, and calculates the gradient of the parameters of each layer.

[0064] 4. Parameter update: Based on the calculated gradient, the computer system uses an optimization algorithm (such as gradient descent, Adam, etc.) to update the internal parameters of the basic water resource allocation algorithm to reduce the loss value.

[0065] The optimization goal of step S14 is to obtain a target water resource allocation algorithm, the degree of fit between the irrigation demand optimization result of the irrigation area template output by the algorithm and the irrigation demand training label is higher than the degree of fit between the irrigation demand estimation information output by the basic water resource allocation algorithm and the training label. The degree of fit is usually measured by calculating the loss function value. The smaller the loss function value, the higher the degree of fit. In order to achieve this goal, the computer system continuously iterates and updates the algorithm parameters during the training process until the loss function value converges to a lower level or reaches a preset number of iterations. During the iteration process, the computer system can also adopt some strategies to accelerate convergence and avoid overfitting, such as learning rate decay, early stopping, regularization, etc.

[0066] In each iteration, the computer system uses a new set of training data (including input features, estimated outputs, and true labels) to update the algorithm parameters. As the iterations proceed, the algorithm gradually learns the complex relationship between the input features and the irrigation demand, and the estimated irrigation demand becomes closer and closer to the true value. In order to accelerate the iteration process and improve the performance of the algorithm, the computer system can use different optimization algorithms such as batch gradient descent, stochastic gradient descent, or mini-batch gradient descent. These algorithms use different amounts of training data in each iteration, and each has its own advantages and disadvantages. For example, batch gradient descent uses all the training data to update the parameters once, which is computationally intensive but converges stably; stochastic gradient descent uses only one training data at a time to update the parameters, which is computationally intensive but may converge more slowly and fluctuate more; mini-batch gradient descent is somewhere in between, using a small portion of the training data each time to update the parameters, which ensures convergence speed and reduces computational complexity. During the training process, the computer system regularly evaluates the performance of the target water resource allocation algorithm. This is usually achieved by calculating indicators such as loss function values ​​or accuracy on a validation set or a test set. If the algorithm performs well on the validation set or test set (i.e., the loss function value is small or the accuracy is high), it can be considered that the algorithm has converged to a good state, and the training can be stopped and the model parameters can be saved.

[0067] In one implementation, the water resource allocation sequence characteristics include water resource allocation variables of the reservoir template at a first number of time points, a first number of irrigation nodes at the first number of time points, and basic irrigation district information of the irrigation district template; step S13, using a basic water resource allocation algorithm, combined with the water resource allocation sequence characteristics and the basin basic characteristics, to estimate the irrigation demand of the irrigation district template, and obtain the irrigation demand estimation information of the irrigation district template, including:

[0068] Step S131: averaging the water resource allocation variables at the first number of time points to obtain an average water resource allocation feature of the reservoir template;

[0069] Step S132: combining the irrigation district basic information of the first number of irrigation nodes and the irrigation district template, generating irrigation district description features describing the first number of irrigation nodes and the irrigation district template;

[0070] Step S133: adopting the basic water resource allocation algorithm, combining the average water resource allocation characteristics, irrigation district description characteristics, water resource allocation sequence characteristics and basin basic characteristics, the irrigation demand of the irrigation district template is estimated to obtain the irrigation demand estimation information of the irrigation district template.

[0071] In step S131, the computer system processes the water resource allocation sequence features. These features include water resource allocation variables of the reservoir template at a first number of time points, such as water storage, outflow, inflow, etc. In order to extract useful information from these time series data, the computer system averages these variables to obtain an average feature that can represent the water resource allocation of the reservoir during the entire time period. For example, suppose there is a reservoir template whose water resource allocation sequence features include 30 consecutive days of water storage data (unit: million cubic meters), as shown below: Water storage sequence = [100, 102, 98, 105, 101, 103, 99, 104, 100, 97, 106, 102, 98, 101, 103, 100, 98, 102, 105, 99, 101, 104, 98, 100, 103, 101, 99, 102, 101]. In step S131, the water storage sequence is averaged to obtain the average water storage: average water storage = (100 + 102 + 98 + .., + 101) / 30 ≈ 101.00 million cubic meters. This average water storage can be used as one of the average water resource allocation characteristics of the reservoir template for subsequent irrigation demand estimation.

[0072] In step S132, the computer system processes the irrigation nodes and irrigation district basic information to generate irrigation district description features that can describe the irrigation district template. These features are intended to capture key information such as the spatial distribution of different irrigation nodes in the irrigation district, irrigation capacity, soil type, crop type, etc. For example, suppose there is an irrigation district template that contains 5 irrigation nodes (labeled as A, B, C, D, E), each of which has its specific irrigation area, soil type and main crops planted. In addition, the irrigation district template also has basic information such as its overall irrigation area, average altitude, average annual precipitation, etc.

[0073] In step S132, the computer system may encode this information into a feature vector or matrix for subsequent processing. For example, one-hot encoding may be used to represent the location of the irrigation node (assuming that the irrigation area is divided into a grid and each node occupies one grid cell):

[0074] At the same time, numerical codes can also be used to represent other basic information of irrigation districts, such as irrigation area, soil type, crop type, etc., and this information can be combined with the irrigation node location code to form a more comprehensive irrigation district description feature vector or matrix.

[0075] In step S133, the computer system will comprehensively utilize the average water resource allocation characteristics, irrigation district description characteristics, water resource allocation sequence characteristics and basin basic characteristics extracted in the previous steps to estimate the irrigation demand for the irrigation district template through the basic water resource allocation algorithm.

[0076] Assume that a neural network model is used as the basic water resource allocation algorithm. This model may contain multiple layers, such as input layer, hidden layer, and output layer. In the input layer, the features extracted in the previous steps need to be concatenated or fused to form a high-dimensional feature vector as the input of the model.

[0077] Before inputting the features into the neural network model, they are appropriately concatenated or fused. For example, the average water resource allocation features (such as average water storage), irrigation area description features (such as irrigation node location coding and irrigation area basic information coding), water resource allocation sequence features (such as water storage sequence) and basin basic features (such as basin area, average annual precipitation) can be concatenated into a long vector:

[0078] Input feature vector = [average water storage, irrigation node location code, irrigation area basic information code, water storage sequence, basin area, average annual precipitation]. Alternatively, feature fusion technology can be used to combine these features at a deeper level. For example, the attention mechanism can be used to assign different weights to different features, thereby highlighting those features that are more important for irrigation demand estimation.

[0079] The hypothetical neural network model consists of an input layer, two hidden layers, and an output layer. The input layer receives the previously concatenated or fused feature vectors; the hidden layer learns the complex relationship between features through nonlinear transformation; and the output layer outputs the estimated value of irrigation demand.

[0080] During the training process, the irrigation demand training labels are used to supervise the learning process of the model. Through optimization methods such as back propagation algorithm and gradient descent, the parameters of the model are continuously adjusted to reduce the difference between the estimated irrigation demand and the actual irrigation demand.

[0081] In the output layer, the irrigation demand estimation results can be processed according to specific needs. For example, if it is necessary to obtain the irrigation demand estimation value of each irrigation node in the irrigation area template, the output layer can be designed as an array of neurons with the same number of irrigation nodes. The output value of each neuron will represent the irrigation demand estimation value of the corresponding irrigation node.

[0082] In addition, if you only need to get an overall irrigation demand estimate (such as the total irrigation demand of the entire irrigation area), the output layer can be designed as a single neuron, and its output value will represent the estimated irrigation demand of the entire irrigation area.

[0083] Through the processing of step S133, the computer system can comprehensively utilize a variety of characteristic information to estimate the irrigation demand for the irrigation area template, providing strong support for subsequent water resource allocation decisions.

[0084] In one implementation, step S133 uses a basic water resource allocation algorithm to estimate the irrigation demand of the irrigation area template by combining the average water resource allocation characteristics, the irrigation area description characteristics, the water resource allocation sequence characteristics, and the basic characteristics of the watershed, and obtains the irrigation demand estimation information of the irrigation area template, including:

[0085] Step S1331: using the feature encoder of the basic water resource allocation algorithm to encode the water resource allocation sequence features to obtain a water resource encoding vector for estimating the irrigation demand area;

[0086] Step S1332: using the integrator of the basic water resource allocation algorithm to integrate the average water resource allocation characteristics and the basin basic characteristics to obtain integrated characteristics;

[0087] Step S1333: using the feature transformer of the basic water resources allocation algorithm to perform feature transformation on the integrated feature to obtain the main integrated feature in the integrated feature;

[0088] Step S1334: using the classifier of the basic water resource allocation algorithm, combined with the water resource coding vector, the main integrated features and the irrigation district description features, the irrigation demand estimation information of the irrigation district template is estimated.

[0089] In step S1331, the computer system processes the water resource allocation sequence features. These features include water resource allocation variables of the reservoir at different time points, such as water storage, outflow, inflow, etc., which can exist in the form of time series. In order to convert these time series data into a format that can be processed by the model and extract useful information from them, the computer system uses the feature encoder in the basic water resource allocation algorithm to encode the water resource allocation sequence features.

[0090] The feature encoder is, for example, a neural network layer or a combination of layers that can learn how to map input data to a new feature space. In this new feature space, useful information in the original data is retained, while noise and redundant information are removed or weakened. In step S1331, the feature encoder may use structures such as a recurrent neural network (RNN), a long short-term memory network (LSTM), or a gated recurrent unit (GRU) to process time series data. These structures can capture the temporal dependencies in the data, that is, the mutual influence between data at different time points.

[0091] Suppose there is a water resources allocation sequence feature, which contains 30 consecutive days of water storage data (unit: million cubic meters), as shown below: Water storage sequence = [100,102,98,105,101,103,99,104,100,97,106,102,98,101,103,100,98,102,105,99,101,104,98,100,103,101,99,102,101].

[0092] In step S1331, the feature encoder receives the water storage sequence as input and processes it through its internal network structure. The processing process includes the following steps:

[0093] 1. Sequence embedding: First, each value in the water storage sequence is converted into a fixed-length vector (called an embedding vector). This embedding process can be achieved through an embedding layer, which maps each value into a high-dimensional vector space.

[0094] 2. Capturing temporal dependencies: Then, recurrent neural networks or similar structures are used to process the embedded sequence data. These structures can capture the temporal dependencies in the data, that is, the mutual influence between the embedded vectors at different time points.

[0095] 3. Encoding vector generation: Finally, the processed sequence data is converted into a fixed-length vector (called encoding vector) through an encoding layer. This encoding vector contains useful information in the original water storage sequence and is used for subsequent irrigation demand estimation.

[0096] After the above processing, a water resource encoding vector for estimating irrigation demand areas is obtained. This vector can be regarded as a high-level representation of the original water storage series, which captures the time dependency and key information in the data.

[0097] In step S1332, the computer system integrates the average water resource allocation characteristics and the basic characteristics of the basin. These characteristics represent the average water resource allocation of the reservoir at different time points and the basic conditions of the basin such as geography, climate, and hydrology. In order to combine these characteristics and extract useful information from them, the computer system uses the integrator in the basic water resource allocation algorithm for processing. The integrator is, for example, a neural network layer or a combination of a group of layers, which can learn how to integrate multiple input features into a unified feature representation. In this unified feature representation, the relevant information between different input features is retained and used for subsequent model predictions. In step S1332, the integrator may use structures such as a fully connected layer (FullyConnectedLayer) or a convolutional layer (ConvolutionalLayer) to process the input features. These structures can learn the complex relationships between input features and integrate them into a higher-level feature representation.

[0098] Assume that there is an average water resource allocation feature (such as average water storage) and a basin basic feature (such as basin area and average annual precipitation), which are expressed as follows: average water storage = 101.00 million cubic meters; basin area = 10,000 square kilometers; average annual precipitation = 800 mm. In step S1332, the integrator receives these features as input and processes them through its internal network structure. The processing process includes the following steps, for example:

[0099] 1. Feature concatenation: First, the average water resource allocation feature and the basin basic feature are concatenated into a long vector, which contains the information of all input features.

[0100] 2. Non-linear transformation: Then, one or more fully connected layers are used to perform non-linear transformations on the concatenated feature vectors. These transformations can learn the complex relationships between the input features and integrate them into a higher-level feature representation.

[0101] 3. Integrated feature generation: Finally, an output layer is used to generate integrated features. This integrated feature is a fixed-length vector that contains useful information from the original input features and is used for subsequent irrigation demand estimation.

[0102] After the above processing, an integrated feature vector is obtained. This vector combines the information of average water resources allocation characteristics and basin basic characteristics, providing strong support for subsequent irrigation demand estimation.

[0103] In step S1333, the computer system further processes the integrated features to extract the main information therein. This main information is crucial for subsequent irrigation demand estimation. To achieve this goal, the computer system uses the feature transformer in the basic water resource allocation algorithm to perform feature transformation (such as mapping) on ​​the integrated features.

[0104] The feature transformer is, for example, a neural network layer or a combination of layers that can learn how to map input features to a new feature space. In this new feature space, the main information in the original features is retained, while the secondary information is removed or weakened. This mapping process helps to extract key information from the input features and improve the prediction performance of the model. In step S1333, the feature transformer may use an autoencoder, principal component analysis (PCA) or similar dimensionality reduction techniques to process the integrated features. These techniques can learn low-dimensional representations of input features and retain the main information therein.

[0105] In step S1334, the water resource coding vector, main integrated features and irrigation area description features are comprehensively used to estimate the irrigation demand of the irrigation area template through the classifier in the basic water resource allocation algorithm. This process is the last step of irrigation demand estimation and is also the most critical step.

[0106] The classifier is, for example, a neural network layer or a combination of layers that can learn how to map input features to output labels or estimates. In irrigation demand estimation, the goal of the classifier is to learn how to map input features (including water resource coding vectors, main integrated features, and irrigation area description features) to irrigation demand estimates. In step S1334, the classifier may use structures such as fully connected layers, softmax layers, or sigmoid layers to process input features. These structures can learn the complex relationship between input features and irrigation demand and output an estimate of irrigation demand.

[0107] Assume that the water resource coding vector, the main integration features and the irrigation area description features have been obtained. In step S1334, the classifier receives these features as input and processes them through its internal network structure. The processing process includes the following steps:

[0108] 1. Feature fusion: First, the water resource coding vector, main integration features and irrigation area description features are fused. This fusion process can be achieved by feature concatenation, feature addition or feature concatenation followed by a fully connected layer. The fused feature vector contains information about all input features.

[0109] 2. Non-linear transformation: The fused feature vector is then transformed non-linearly using one or more fully connected layers. These transformations are able to learn the complex relationships between input features and irrigation demand and map them to an estimate of irrigation demand.

[0110] 3. Generation of irrigation demand estimation information: Finally, an output layer is used to generate irrigation demand estimation information. This information may be a numerical value (indicating the total amount or intensity of irrigation demand) or a probability distribution (indicating the possibility of different irrigation demands).

[0111] In one implementation, step S1331 uses a feature encoder of a basic water resource allocation algorithm to encode the water resource allocation sequence features to obtain a water resource encoding vector for estimating the irrigation demand area, including:

[0112] Step S13311: using the feature encoder of the basic water resource allocation algorithm to generate a sequence encoding representation describing the features of the water resource allocation sequence;

[0113] Step S13312: Incorporating distribution information of the first number of irrigation nodes into the sequence code representation to obtain a temporary code representation;

[0114] Step S13313: performing significance weighting on the temporary coding representation based on the first significance factor, the second significance factor and the third significance factor of the feature encoder respectively, to obtain a first candidate coding representation, a second candidate coding representation and a third candidate coding representation;

[0115] Step S13314: Determine a water resource coding vector for estimating the irrigation demand area by combining the first candidate coding representation, the second candidate coding representation, and the third candidate coding representation.

[0116] In step S13311, the computer system encodes the water resource allocation sequence features using a feature encoder to generate a sequence code representation that can describe these features. The water resource allocation sequence features usually include water resource allocation variables at a series of time points, such as water storage, outflow, inflow, etc., which exist in the form of time series.

[0117] The feature encoder can be a recurrent neural network (RNN), a long short-term memory network (LSTM), a gated recurrent unit (GRU), or other neural network structures that can process sequence data. These structures can capture the temporal dependencies in the sequence data, that is, the mutual influence between data at different time points. During the encoding process, the feature encoder receives the water resource allocation sequence features as input and processes them through its internal network structure. The processing process includes, for example, steps such as sequence embedding, temporal dependency capture, and encoding vector generation. Finally, the feature encoder outputs a fixed-length vector as a sequence encoding representation of the water resource allocation sequence features.

[0118] For example, suppose there is a water resource allocation sequence feature, which contains water storage data for 7 consecutive days (unit: million cubic meters), as shown below: Water storage sequence = [100, 105, 102, 108, 104, 101, 106]. In step S13311, the feature encoder receives this water storage sequence as input and processes it through its internal network structure. The processing process includes, for example: First, each value in the water storage sequence is converted into a fixed-length vector (called an embedding vector). This embedding process can be implemented by an embedding layer (EmbeddingLayer), which maps each value to a high-dimensional vector space. For example, each water storage value can be mapped to a vector space of length 10 to obtain a 7×10 embedding matrix. Then, a recurrent neural network or similar structure is used to process the embedded sequence data. These structures can capture the temporal dependency in the data, that is, the mutual influence between the embedded vectors at different time points. For example, an LSTM network can be used to process the embedding matrix to obtain a fixed-length hidden state vector as a sequence encoding representation. After the above processing, a sequence encoding representation describing the characteristics of the water resources allocation sequence is obtained. It is a vector of fixed length that contains useful information in the original sequence data.

[0119] In step S13312, the computer system incorporates the distribution information of the irrigation nodes into the sequence code representation. Irrigation nodes are key locations within the irrigation area, and they determine the allocation and utilization of water resources. By incorporating the distribution information of the irrigation nodes, the irrigation demand area can be more accurately estimated.

[0120] The distribution information of irrigation nodes can be represented in a variety of ways, such as spatial coordinates, grid index, one-hot encoding, etc. In step S13312, these representations can be converted into a format compatible with the sequence coding representation for subsequent fusion processing. The fusion process can be achieved by feature concatenation, feature addition, feature concatenation followed by a fully connected layer, etc. In step S13312, a suitable method can be selected to integrate the distribution information of irrigation nodes into the sequence coding representation. For example, suppose there is an irrigation district template that contains 4 irrigation nodes (labeled as A, B, C, D), each of which has its specific spatial coordinates. In step S13312, these spatial coordinates can be converted into a format compatible with the sequence coding representation (such as one-hot encoding or embedded vector) and fused with the sequence coding representation. For example, the spatial coordinates of each irrigation node can be converted into a one-hot encoding vector of length 4 (assuming that the irrigation district is divided into a 4×4 grid), resulting in a 4×4 one-hot encoding matrix. Then, this one-hot encoding matrix can be concatenated with the sequence encoding representation to obtain a temporary encoding representation containing the irrigation node distribution information.

[0121] In step S13313, the computer system uses the significance factor of the feature encoder to weight the temporary encoding representation to highlight the important information and suppress the secondary information. This weighting process helps to improve the accuracy and robustness of the encoding vector.

[0122] Saliency factors can be understood as attention weights, which represent the importance of different parts in the temporary encoding representation. In the feature encoder, these saliency factors can be learned and used to guide subsequent weighting processing.

[0123] The weighting process can be implemented by element-by-element multiplication, weighted summation, etc. In step S13313, a suitable method can be selected to apply the significance factor to weight the temporary coding representation. For example, suppose a temporary coding representation of length 20 (denoted as H) and three significance factors (denoted as α1, α2, α3) are obtained, all of which are vectors of length 20. In step S13313, these significance factors can be used to weight the temporary coding representation to obtain three candidate coding representations (denoted as H1, H2, H3): H1=H*α1; H2=H*α2; H3=H*α3; where * represents an element-by-element multiplication operation. After the above weighted processing, three candidate coding representations are obtained, which respectively highlight the importance of different parts of the temporary coding representation.

[0124] In step S13314, the computer system combines the three candidate coding representations to determine the final water resource coding vector. This coding vector will be used in the subsequent irrigation demand estimation process. The combination method can be achieved by feature concatenation, feature addition, feature concatenation followed by a fully connected layer, etc. In step S13314, a suitable method can be selected to combine the three candidate coding representations. For example, assume that three candidate coding representations of length 20 are obtained (denoted as H1, H2, H3). In step S13314, you can choose to concatenate them to obtain a vector of length 60 (denoted as F): F=[H1,H2,H3]; then, a fully connected layer can be used to perform a nonlinear transformation on the concatenated vector to obtain a fixed-length water resource coding vector (denoted as E): E=W*F+b; where W is a weight matrix and b is a bias vector. After the above processing, a water resource coding vector E for estimating the irrigation demand area is obtained. This vector contains useful information about the characteristics of the water resource allocation sequence and the distribution information of irrigation nodes, and is used in the subsequent irrigation demand estimation process.

[0125] In one implementation, step S13314, combining the first candidate coding representation, the second candidate coding representation, and the third candidate coding representation to determine a water resource coding vector for estimating the irrigation demand area, includes:

[0126] Step S133141: multiplying the first candidate coding representation and the second candidate coding representation to obtain a significance vector describing that the irrigation area template includes the irrigation demand area;

[0127] Step S133142: Combine the product between the significance vector and the third candidate coding representation to obtain a water resource coding vector for estimating the irrigation demand area.

[0128] In step S133141, the goal of the computer system is to generate a vector that can describe the significance of the irrigation demand area in the irrigation area template. This vector will help to more accurately estimate the irrigation demand area in the subsequent steps. To achieve this goal, the computer system performs a multiplication operation on the first candidate code representation and the second candidate code representation. The multiplication of two vectors can be performed by element-by-element multiplication (Hadamard product), because it can retain the importance of each element in the vector and highlight those elements that are more significant in both the first candidate code representation and the second candidate code representation through the multiplication operation. Assume that the first candidate code representation is vector H1 and the second candidate code representation is vector H2, both of which are vectors of length n. Then, the element-by-element multiplication operation can be expressed as: significance vector = H1*H2; where * represents element-by-element multiplication. The result of the multiplication is a new vector (denoted as S), each of whose elements is the product of the corresponding elements of H1 and H2.

[0129] Each element in the significance vector S reflects the significance of the corresponding position in the irrigation demand area within the irrigation area template. If the elements of H1 and H2 at a certain position are both large, then the result after multiplication will also be large, indicating that the position has a higher significance in the irrigation demand area. On the contrary, if the elements of H1 and H2 at a certain position are both small or one is larger and the other is smaller, then the result after multiplication will also be small, indicating that the significance of the position in the irrigation demand area is low. For example, suppose there are the following two candidate encoding representation vectors (both with a length of 4, just for example): H1=[0.8,0.2,0.5,0.9]; H2=[0.6,0.9,0.3,0.7]; these two vectors are multiplied element by element to obtain: S=[0.8*0.6,0.2*0.9,0.5*0.3,0.9*0.7]=[0.48,0.18,0.15,0.63].

[0130] This significance vector S represents the significance of different positions in the irrigation area template in the irrigation demand area. For example, the significance of the first position is 0.48, indicating that the position has a high significance in the irrigation demand area, while the significance of the third position is 0.15, indicating that the position has a low significance in the irrigation demand area.

[0131] In step S133142, the computer system aims to combine the saliency vector with the third candidate coding representation to generate a final water resource coding vector. This vector will be used in the subsequent irrigation demand estimation process. The combination process can be achieved by element-by-element multiplication, weighted summation, or feature concatenation followed by a fully connected layer. In this particular context, element-by-element multiplication is a simple and effective method because it can retain the importance of each element in the saliency vector and the third candidate coding representation, and further highlight those elements that are significant in the irrigation demand area through the multiplication operation. Assume that the saliency vector is S and the third candidate coding representation is vector H3, both of which are vectors of length n. Then, the element-by-element multiplication operation can be expressed as: water resource coding vector = S*H3; where * represents element-by-element multiplication. The result of the multiplication is a new vector (denoted as E), each element of which is the product of the corresponding elements of S and H3.

[0132] Each element in the water resource coding vector E reflects the importance of the corresponding position in the irrigation demand area within the irrigation area template and the water resource allocation of the position. Since E is the element-by-element product of the saliency vector S and the third candidate coding representation H3, it combines the saliency of the irrigation demand area and the water resource allocation information, providing strong support for subsequent irrigation demand estimation. For example, continuing the above example, assume that there is the following third candidate coding representation vector (length is also 4): H3=[0.7,0.5,0.8,0.6]; now, the saliency vector S is multiplied element-by-element by H3 to obtain the water resource coding vector E: E=[0.48*0.7,0.18*0.5,0.15*0.8,0.63*0.6]=[0.336,0.09,0.12,0.378].

[0133] This water resource coding vector E represents the water resource allocation and its importance at different locations in the irrigation demand area within the irrigation area template. For example, the water resource coding of the first location is 0.336, indicating that the location has a high importance and significance of water resource allocation in the irrigation demand area; while the water resource coding of the third location is 0.12, indicating that the location has a low importance and significance of water resource allocation in the irrigation demand area.

[0134] In one implementation, step S1334 uses a classifier of a basic water resource allocation algorithm, combines water resource coding vectors, main integration features, and irrigation district description features, and estimates irrigation demand for the irrigation district template to obtain irrigation demand estimation information for the irrigation district template, including:

[0135] Step S13341: using the classifier of the basic water resource allocation algorithm, the irrigation area description feature is divided to obtain a second number of sub-irrigation area description features; one or more sub-irrigation area description features of the second number of sub-irrigation area description features include irrigation nodes, and do not include irrigation areas of the irrigation area template;

[0136] Step S13342: performing feature fusion on the water resource coding vector, the main integration feature, and the second number of sub-irrigation area description features to obtain fusion features;

[0137] Step S13343: Use the classifier in combination with the fusion features to estimate the irrigation demand for the irrigation area template, and obtain the irrigation demand estimation information of the irrigation area template.

[0138] In step S13341, the goal of the computer system is to divide the irrigation area description features to obtain more fine-grained sub-irrigation area description features. These sub-irrigation area description features will help to more accurately describe different areas within the irrigation area and provide more valuable information for subsequent irrigation demand estimation. The irrigation area description features usually include information on the geographical location, soil type, crop type, irrigation facilities, etc. of the irrigation area. In order to obtain more fine-grained sub-irrigation area description features, the computer system can adopt a variety of division methods, such as grid division based on geographical location, regional division based on crop type, or division based on the coverage of irrigation facilities.

[0139] Assuming that the irrigation area is divided into an MxN grid (for example, a 10x10 grid) based on geographic location, each grid can be regarded as a sub-irrigation area with its own unique irrigation area description features. These sub-irrigation area description features may include the geographic location of the sub-irrigation area (such as the row and column index of the grid), soil type, crop type, etc. For example, suppose there is an irrigation area template whose irrigation area description features contain the following information:

[0140] Geographic location: The irrigation area is located between 30° and 31° north latitude and 110° and 111° east longitude. Soil type: mainly clay and sand. Crop type: mainly wheat and corn. The irrigation area is divided into a 10x10 grid using a geographic location-based grid division method. After division, 100 sub-irrigation areas are obtained (each grid is a sub-irrigation area), and each sub-irrigation area has its own unique irrigation area description features. For example, the sub-irrigation area located in the 5th row and the 3rd column may have the following irrigation area description features: Geographic location: 30.4° north latitude, 110.6° east longitude (the center point coordinates of the grid). Soil type: clay. Crop type: wheat. In this way, the original irrigation area description features are successfully divided into more fine-grained sub-irrigation area description features, providing richer information for subsequent feature fusion and irrigation demand estimation.

[0141] In step S13342, the computer system aims to fuse the water resource coding vector, the main integration feature, and the second number of sub-irrigation area description features to obtain a fused feature containing more comprehensive information. This fused feature will serve as input for subsequent irrigation demand estimation.

[0142] Feature fusion can be achieved in many ways, such as feature concatenation, feature addition, feature concatenation followed by a fully connected layer, etc. In this specific context, feature concatenation can be chosen to perform feature fusion.

[0143] In step S13343, the computer system aims to process the fused features using the classifier to obtain the irrigation demand estimation information of the irrigation area template. This estimation information will help understand the irrigation demand of different areas in the irrigation area template.

[0144] A classifier can be a neural network layer or a combination of layers that learns how to map input features (in this case, fused features G) to output labels or estimates. In irrigation demand estimation, the goal of the classifier is to learn how to map fused features to estimates of irrigation demand. Assume that a fully connected neural network is used as a classifier, which consists of an input layer, a hidden layer, and an output layer. The input layer receives the fused features G as input; the hidden layer learns the complex relationship between the input features and irrigation demand through nonlinear transformation; and the output layer outputs an estimate of irrigation demand.

[0145] In the calculation process of irrigation demand estimation, the classifier receives the fusion feature G as input and processes it through its internal network structure. The specific method can be referred to the above description.

[0146] In one implementation, step S14, combining the irrigation demand training label and the irrigation demand estimation information, training the basic water resource allocation algorithm to obtain the target water resource allocation algorithm, including:

[0147] Step S141: combining the irrigation demand training label and the irrigation demand estimation information to determine the irrigation demand area estimation cost of the basic water resource allocation algorithm;

[0148] Step S142: Determine the stability of the basic water resource allocation algorithm in combination with the estimated cost of the irrigation demand area;

[0149] Step S143: When the basic water resource allocation algorithm is unstable, the basic water resource allocation algorithm is corrected in combination with the estimated cost of the irrigation demand area to obtain a target water resource allocation algorithm.

[0150] In step S141, the computer system determines the estimated cost of the irrigation demand area of ​​the basic water resource allocation algorithm. This cost measures the difference between the irrigation demand area estimated by the algorithm and the actual irrigation demand area, and is an important indicator for evaluating the performance of the algorithm.

[0151] The estimated cost of the irrigation demand area can be obtained by calculating a certain distance or difference between the estimated irrigation demand area and the actual irrigation demand area. Common distance or difference measurement methods include mean square error (MSE), cross entropy loss, IoU (intersection over union), etc.

[0152] Assuming that mean square error (MSE) is used as a measure of the estimated cost of the irrigation demand area, the calculation formula is: ; where N is the number of samples, y i is the actual irrigation requirement area of ​​the ith sample (can be a binary image or vector), is the estimated irrigation demand area of ​​the ith sample, and MSE represents the average of the sum of squares of the differences between the true values ​​and the estimated values ​​of all samples.

[0153] For example, suppose you have the following real irrigation demand labels and irrigation demand estimates:

[0154] True irrigation demand labels (binarized images): [[0,0,1,1],[0,1,1,0],[1,1,0,0],[0,0,0,0]]. 1 indicates irrigation demand and 0 indicates no irrigation demand.

[0155] Irrigation demand estimation information (binarized image): [[0,0,1,1],[0,1,0,0],[1,1,1,0],[0,0,0,0]], Now, calculate the mean square error (MSE) between these two binary images:

[0156] ; This MSE value measures the difference between the estimated irrigation demand information and the actual irrigation demand label.

[0157] In step S142, the computer system evaluates the stability of the basic water resource allocation algorithm in combination with the estimated cost of the irrigation demand area. The stability of the algorithm refers to the consistency of its performance on different data sets or different training rounds. A stable algorithm can maintain relatively consistent performance under different circumstances.

[0158] The stability of an algorithm can be evaluated in many ways, such as calculating the performance difference of the algorithm on different data sets, observing the change in the loss function value of the algorithm during training, etc. A common method is to calculate the standard deviation or coefficient of variation of the performance of the algorithm on different data sets to measure its stability.

[0159] Assuming that the performance standard deviation is used as a measure of stability, the calculation formula is: ; where K is the number of different data sets or different training rounds, P i is the performance of the algorithm on the i-th data set or training round (such as accuracy, recall, etc.), It is the average performance over all datasets or training rounds.

[0160] This standard deviation value measures the fluctuation of the algorithm's performance on different data sets. The smaller the standard deviation, the more stable the algorithm.

[0161] In step S143, when the computer system detects that the basic water resource allocation algorithm is unstable (i.e., the performance of the algorithm on different data sets or different training rounds fluctuates greatly), it needs to revise the algorithm to improve its stability and reduce the estimated cost of the irrigation demand area.

[0162] The modification of the algorithm can be achieved in many ways, such as adjusting the hyperparameters of the algorithm (such as learning rate, batch size, regularization strength, etc.), changing the structure of the algorithm (such as increasing or decreasing the number of network layers, changing the activation function, etc.), introducing regularization techniques (such as L1 regularization, L2 regularization, Dropout, etc.), etc. Assume that adjusting the learning rate of the algorithm and introducing Dropout regularization are selected as the modification methods. The adjustment of the learning rate can affect the convergence speed and stability of the algorithm during the training process; Dropout regularization prevents overfitting by randomly discarding a part of neurons during the training process, thereby improving the generalization ability of the algorithm. For example, assuming that the instability of the basic water resource allocation algorithm is high (such as a large standard deviation) in step S142, it is now selected to reduce the learning rate from 0.01 to 0.001, and introduce Dropout regularization after the fully connected layer (the discard rate is 0.5). Then, retrain the algorithm and evaluate its stability and the estimated cost of the irrigation demand area.

[0163] After correction, the following results may be obtained: new irrigation demand area estimation cost (MSE): 2 / 16 (lower than the previous 3 / 16), new algorithm stability (standard deviation): 0.01 (lower than the previous 0.016), which shows that by adjusting the learning rate and introducing Dropout regularization, the stability of the basic water resources allocation algorithm and the accuracy of irrigation demand estimation have been successfully improved, thereby obtaining a more robust and accurate target water resources allocation algorithm.

[0164] In one implementation, in step S143, when the basic water resource allocation algorithm is unstable, the basic water resource allocation algorithm is modified in combination with the estimated cost of the irrigation demand area to obtain a target water resource allocation algorithm, including:

[0165] Step S1431: when the basic water resource allocation algorithm is unstable, the algorithm configuration parameters of the basic water resource allocation algorithm are optimized in combination with the estimated cost of the irrigation demand area to obtain an optimized basic water resource allocation algorithm;

[0166] Step S1432: using the optimized basic water resource allocation algorithm, combined with the water resource allocation sequence characteristics and the basic characteristics of the watershed, the irrigation demand of the irrigation area template is estimated to obtain the irrigation demand optimization result of the irrigation area template; the irrigation demand optimization result is the optimization result representing the irrigation demand area included in the irrigation area template;

[0167] Step S1433: combining the irrigation demand optimization result and the irrigation demand training label, determining whether the optimized basic water resource allocation algorithm is stable;

[0168] Step S1434: When the optimized basic water resource allocation algorithm is stable, the optimized basic water resource allocation algorithm is used as the target water resource allocation algorithm.

[0169] In step S1431, the computer system optimizes the algorithm configuration parameters of the basic water resource allocation algorithm. These parameters include, for example, learning rate, batch size, number of network layers, number of neurons, activation function type, regularization strength, etc., which have an important impact on the performance and stability of the algorithm.

[0170] The optimization of algorithm configuration parameters can be performed through a variety of methods, such as grid search, random search, Bayesian optimization, etc. The goal of these methods is to find a set of optimal parameter combinations in a given parameter space so that the algorithm performs best on the training set and validation set while maintaining good generalization ability.

[0171] Suppose that a grid search method is used for parameter optimization. First, a parameter grid is defined, which contains all the parameters to be optimized and their possible value ranges. Then, each parameter combination in the parameter grid is traversed, and these parameter combinations are used to train the algorithm, and its performance is evaluated on the validation set. Finally, the parameter combination with the best performance is selected as the optimal parameter. For example, the basic water resource allocation algorithm is assumed to be a neural network model, and the parameters that need to be optimized include learning rate, batch size, and L2 regularization strength. Now, use the grid search method to traverse each parameter combination in the parameter grid and evaluate the performance of the algorithm on the validation set. Suppose that the optimal parameter combination is found: learning rate = 0.005, batch size = 64, l2 regularization = 0.001. Use this set of optimal parameters to retrain the algorithm and obtain the optimized basic water resource allocation algorithm.

[0172] In step S1432, the computer system uses the optimized basic water resource allocation algorithm to estimate the irrigation demand of the irrigation area template. The input of this process includes the water resource allocation sequence characteristics and the basic characteristics of the watershed, and the output is the irrigation demand optimization result of the irrigation area template.

[0173] The implementation method of irrigation demand estimation depends on the specific structure of the optimized basic water resource allocation algorithm. If the algorithm is a neural network model, the estimation process usually includes two steps: forward propagation and loss calculation. In the forward propagation process, the algorithm receives input features and calculates the output layer by layer through the network layer; in the loss calculation process, the algorithm calculates the difference between the output value and the true label (i.e., the loss value), and uses the loss value to guide the subsequent training and optimization of the algorithm. For example, continuing the above example, assume that the optimized basic water resource allocation algorithm is a neural network model with multiple fully connected layers. Now, use this algorithm to estimate the irrigation demand for the irrigation district template. The water resource allocation sequence features and basin basic features of the irrigation district template are extracted as input features, and the estimated value of irrigation demand is calculated through the forward propagation process. Then, the difference between the estimated value and the true irrigation demand label (i.e., the loss value) is calculated, and the loss value is used to evaluate the performance of the algorithm.

[0174] In step S1433, the computer system evaluates the stability of the optimized basic water resource allocation algorithm. The evaluation of stability is usually based on whether the performance of the algorithm is consistent on different data sets or different training rounds.

[0175] The stability of an algorithm can be evaluated by calculating the variance or standard deviation of the algorithm's performance on different data sets or different training rounds. The smaller the variance or standard deviation, the more stable the algorithm's performance. In addition, methods such as cross-validation can be used to further verify the stability of the algorithm.

[0176] For example, continuing the example above, suppose that multiple training and testing are performed on the optimized basic water resource allocation algorithm, and the performance of each test (such as accuracy, recall, etc.) is recorded. Now, calculate the variance or standard deviation of these performances to evaluate the stability of the algorithm. Assume that a small variance value (such as 0.001) is obtained, which indicates that the performance of the optimized algorithm on different test sets is relatively consistent, so the algorithm can be considered stable.

[0177] In step S1434, when the computer system confirms that the optimized basic water resource allocation algorithm is stable and its performance meets the requirements, it saves and uses the algorithm as the target water resource allocation algorithm.

[0178] The determination of the target water resource allocation algorithm is usually based on whether its performance on the validation set or test set is better than other candidate algorithms and whether its performance is stable. In addition, factors such as the algorithm's computational complexity and operating efficiency also need to be considered.

[0179] For example, continuing the above example, assume that the accuracy of the optimized basic water resource allocation algorithm on the validation set reaches more than 95%, and its performance remains consistent in multiple tests (variance is less than 0.01). In addition, the computational complexity and operating efficiency of the algorithm also meet the requirements. Therefore, the algorithm can be saved and used as the target water resource allocation algorithm. Through the processing of the above four sub-steps, the computer system successfully optimized and evaluated the stability of the basic water resource allocation algorithm, and obtained a more stable and accurate target water resource allocation algorithm. This algorithm will provide strong support for subsequent water resource allocation decisions.

[0180] In one implementation, the method further includes:

[0181] Step S140: obtaining estimated information of a second number of demand areas corresponding to the target irrigation area, where the estimated information of the second number of demand areas is determined in combination with different allocation detail information corresponding to the target irrigation area;

[0182] Step S150: Determine a support coefficient describing the target irrigation area including the irrigation demand area in combination with the estimated information of the second number of demand areas;

[0183] Step S160: when the support coefficient is higher than the support coefficient, obtaining an irrigation network including the target irrigation area;

[0184] Step S170: combining the second number of demand area estimation information, marking the irrigation demand areas of the target irrigation district in the irrigation network.

[0185] In step S140, the computer system aims to obtain multiple demand area estimation information related to the target irrigation area. The estimation information is determined based on the allocation details of the target irrigation area under different time periods or different conditions. By collecting multiple estimation information, the computer system can have a more comprehensive understanding of the irrigation demand of the target irrigation area.

[0186] The demand area estimation information can be obtained by running the target water resource allocation algorithm multiple times, each time using different allocation details as input. Such allocation details include, for example, different water resource allocation sequence characteristics, basin basic characteristics, or other relevant factors. For example, the computer system can generate different allocation details based on historical data, real-time data, or forecast data, and run the algorithm separately to estimate irrigation demand.

[0187] For example, suppose there is a target irrigation district A, and you want to obtain the estimated irrigation demand information of the irrigation district in different time periods. You can collect the allocation details of irrigation district A for each month in the past year (such as monthly water storage, outflow, inflow, etc.), and then run the target water resource allocation algorithm to estimate the irrigation demand. In this way, you can get the estimated information of 12 demand areas, each of which represents the irrigation demand of irrigation district A in the corresponding month.

[0188] In step S150, the task of the computer system is to determine a support coefficient based on the estimated information of multiple demand areas, which is used to describe the confidence or probability that the target irrigation area contains the irrigation demand area. The level of the support coefficient will directly affect the decision-making of subsequent steps.

[0189] The support coefficient can be obtained by calculating the average, median, mode or other statistics of the estimated information of multiple demand areas. In addition, more complex models or methods can be used to integrate multiple estimated information, such as using Bayesian networks, ensemble learning or deep learning techniques to estimate the support coefficient.

[0190] In step S160, the computer system makes a decision based on the support coefficient. If the support coefficient is higher than a certain preset threshold (referred to as the "support coefficient"), the system will consider that the confidence that the target irrigation area contains the irrigation demand area is high enough to warrant further analysis. At this point, the system will obtain the irrigation network that includes the target irrigation area.

[0191] The irrigation network is a network diagram that describes the physical connection relationship between irrigation facilities, pipelines, channels, etc. within the irrigation area. It can help the computer system to more intuitively understand the structural layout and irrigation process of the irrigation area. The irrigation network can be obtained through geographic information system (GIS), remote sensing technology, field surveys, etc. In practical applications, the acquisition of the irrigation network is carried out in cooperation with relevant departments or institutions, such as the Water Conservancy Bureau, the Agricultural Bureau, etc. These institutions usually have detailed irrigation area maps and irrigation network data.

[0192] For example, continuing the above example, assume that the threshold of the support coefficient is set to 0.7 (i.e., 70%). If the support coefficient of irrigation area A in area B reaches 0.83 (higher than the threshold), then the computer system will consider that the confidence that area B has a high irrigation demand is high enough. At this time, the system will try to obtain an irrigation network diagram including irrigation area A in order to further analyze the irrigation demand and resource allocation of area B. In step S170, the final task of the computer system is to mark the irrigation demand area of ​​the target irrigation area in the irrigation network. This step will be implemented based on the previously obtained multiple demand area estimation information and irrigation network diagram. The marking of the irrigation demand area can be achieved by adding marks, color coding, symbols, etc. to the irrigation network diagram. The marked content includes, for example, the intensity, type, time, and other information of the irrigation demand. In actual operation, the computer system uses, for example, GIS software or other graphics processing tools to complete the marking work. During the marking process, the computer system considers the consistency and reliability of the estimation information of multiple demand areas. If multiple estimation information indicates that a certain area has a high irrigation demand, then this area should be marked out. At the same time, the actual conditions and constraints of the irrigation network, such as pipeline capacity and water supply, also need to be considered.

[0193] For example, continuing with the above example, assume that an irrigation network diagram including irrigation area A has been obtained. Now, you want to mark the irrigation demand of irrigation area A in area B on the diagram. Based on the 12-month irrigation demand estimate information obtained previously, the intensity and type of irrigation demand in area B can be determined (such as high frequency, large amount of irrigation, etc.). Then, use GIS software to add corresponding marks or color codes on the irrigation network diagram to represent this information. For example, you can use red areas to represent areas with high irrigation demand, and mark the intensity and type of irrigation demand next to them.

[0194] In one implementation, the method provided by the embodiment of the present invention may further include:

[0195] Step S180: obtaining strategy reference data of the target reservoir, the strategy reference data including basin basic characteristics, a first strategy tag and a second strategy tag, the first strategy tag and the second strategy tag are determined according to the scheduling strategy of the target reservoir for irrigation demand in different monitoring cycles;

[0196] Step S190: Generate strategy description data corresponding to the target reservoir according to the basic characteristics of the watershed, the first strategy mark and the second strategy mark;

[0197] Step S200: extracting strategy marker features from strategy description data through a strategy marker prediction algorithm to obtain extracted strategy marker features, and performing strategy marker decoding on the extracted strategy marker features to obtain a predicted strategy marker of a target reservoir.

[0198] In step S180, the computer system obtains the strategy reference data of the target reservoir. These data are important bases for formulating water resources allocation strategies, and they include information such as basic characteristics of the basin, the first strategy mark and the second strategy mark.

[0199] The basic characteristics of the watershed describe the basic conditions of the geography, climate, and hydrology of the watershed where the target reservoir is located, such as watershed area, average annual precipitation, river length, soil type, etc. These characteristics provide basic information for formulating scheduling strategies.

[0200] The first strategy marker and the second strategy marker are determined based on the scheduling strategy of the target reservoir for irrigation demand in different monitoring periods. They represent the response strategy and action guide of the reservoir under different conditions (such as drought, flood, etc.). For example, the first strategy marker may indicate reducing non-agricultural water use to ensure agricultural irrigation under drought conditions, while the second strategy marker may indicate increasing reservoir flood discharge to relieve downstream pressure under flood conditions.

[0201] Strategy reference data can be obtained through a variety of channels, such as historical records, expert consultation, on-site research, etc. In a computer system, this data can exist in the form of a database, file, or API interface, which is convenient for the computer system to read and process.

[0202] For example, suppose there is a target reservoir C, whose basin has a drainage area of ​​1,000 square kilometers and an average annual precipitation of 800 mm. In the past few years, the reservoir has adopted a strategy of reducing non-agricultural water use under drought conditions (first strategy marker) and a strategy of increasing reservoir flood discharge under flood conditions (second strategy marker). Now, the computer system reads these strategy reference data from the relevant database or file for subsequent processing.

[0203] In step S190, the computer system generates the strategy description data corresponding to the target reservoir based on the strategy reference data obtained in step S180. These data will be used in the subsequent strategy tag feature extraction and decoding process. The strategy description data is, for example, a structured data set that includes the basic characteristics of the watershed, the first strategy tag, the second strategy tag, and other related information. These data can be stored in a table, a JSON object, or other format. When generating the strategy description data, the computer system associates the basic characteristics of the watershed with the strategy tag, and performs some preprocessing operations, such as data cleaning, normalization, etc.

[0204] In step S200, the computer system processes the strategy description data generated in step S190 using a strategy tag prediction method to obtain a predicted strategy tag of the target reservoir. This process includes two sub-steps: strategy tag feature extraction and strategy tag decoding.

[0205] The strategy tag feature extraction is the first step of the strategy tag estimation algorithm, which aims to extract useful feature information from the strategy description data. These features include, for example, numerical representations of basin basic features, text representations of strategy tags, or other related information. When extracting features, computer systems can use various techniques, such as word embedding, feature selection, dimensionality reduction, etc.

[0206] Strategy tag decoding is the second step of the strategy tag prediction algorithm, which aims to convert the extracted strategy tag features into understandable prediction strategy tags. This process usually involves classification, regression or other machine learning techniques. The result of decoding will be one or more prediction strategy tags, which represent the computer system's prediction of the future operation strategy of the target reservoir.

[0207] The policy tag prediction algorithm can be a complex machine learning model, such as a deep learning network, random forest, support vector machine, etc. These models can learn the complex relationship between policy description data and predicted policy tags, and generate accurate prediction results based on new policy description data.

[0208] In practical applications, policy tag prediction algorithms, for example, undergo extensive training and optimization to achieve ideal performance. This usually involves training the model using historical data and adjusting model parameters through methods such as cross-validation and grid search.

[0209] In one implementation, step S200 extracts strategy tag features from strategy description data by using a strategy tag prediction algorithm to obtain extracted strategy tag features, and performs strategy tag decoding on the extracted strategy tag features to obtain a predicted strategy tag of a target reservoir, including:

[0210] Step S201: extracting the strategy tag feature of each description item included in the strategy description data through a tag feature encoder of a strategy tag estimation algorithm, and obtaining a first strategy tag feature corresponding to each description item;

[0211] Step S202: performing strategy tag feature learning on the first strategy tag feature corresponding to each description item through the learner of the strategy tag estimation algorithm to obtain the second strategy tag feature corresponding to each description item;

[0212] Step S203: down-sampling the second strategy tag feature corresponding to each description item by using a strategy tag estimation algorithm to obtain a tag sampling feature corresponding to the strategy description data;

[0213] Step S204: classify the label sampling features corresponding to the strategy description data through the strategy classifier of the strategy label prediction algorithm to obtain the predicted strategy label of the target reservoir.

[0214] In step S201, the task of the computer system is to use the tag feature encoder of the strategy tag estimation algorithm to extract features from each description item in the strategy description data. These description items include, for example, basic features of the watershed (such as watershed area, average annual precipitation, etc.), historical strategy tags (such as the first strategy tag, the second strategy tag, etc.) and other related information.

[0215] A labeled feature encoder is, for example, a neural network layer or a combination of layers that learns how to map input data (in this case, the description items of the policy description data) to a new feature space where the useful information in the original data is preserved and can be more easily processed by subsequent machine learning models.

[0216] In practical applications, the tag feature encoder may adopt a convolutional neural network (CNN), a recurrent neural network (RNN), a long short-term memory network (LSTM), a gated recurrent unit (GRU), or a Transformer structure, etc. These structures can capture local features, temporal dependencies, or global dependencies in the input data.

[0217] In step S201, the marker feature encoder receives the data item as input and processes it through its internal network structure. The processing process includes the following steps:

[0218] 1. Embedding: First, convert the two numerical features, drainage area and average annual precipitation, into embedding vectors. This can be achieved through a lookup table or a fully connected layer. For example, you can create an embedding matrix for each feature and retrieve the corresponding embedding vector from the embedding matrix based on the numerical index of the feature.

[0219] 2. Encoding: Then, the embedding vector is encoded using structures such as convolutional layers, recurrent layers, or Transformer layers. The encoding process aims to extract the complex relationships between features and map them into a new feature space. For example, the embedding vector can be convolved using a convolutional layer to capture local features, or the embedding vector can be processed sequentially using an LSTM layer to capture temporal dependencies.

[0220] 3. Output: Finally, the encoder outputs a fixed-length vector as the first strategy tag feature. This vector contains useful information about the input data item and is used for subsequent machine learning processing.

[0221] In step S202, the task of the computer system is to further learn the first strategy tag feature corresponding to each description item using the learner of the strategy tag prediction algorithm. This process aims to extract deeper feature information and provide more valuable input for subsequent classification tasks.

[0222] A learner is, for example, a neural network layer or a combination of layers that learns how to map input features (in this case, first policy labeled features) to higher-level feature representations. This process may involve operations such as nonlinear transformations, feature interactions, or feature selection.

[0223] In practical applications, learners may use structures such as fully connected layers, convolutional layers, pooling layers, attention mechanisms, or graph neural networks. These structures can capture the complex relationships between input features and integrate them into a more compact feature representation.

[0224] For example, continuing the above example, assume that the first strategy marker features corresponding to the two descriptive items of watershed area and average annual precipitation have been obtained. Now, these features are input into the learner for further learning. The processing of the learner includes, for example: First, a nonlinear transformation of the first strategy marker features is performed using a fully connected layer. This can be achieved by applying activation functions such as ReLU, sigmoid, or tanh to increase the expressive power of the model. Then, structures such as attention mechanisms or graph neural networks are used to capture the interactive relationship between features. For example, attention mechanisms can be used to assign different weights to different features to highlight those features that are more important for the classification task; or graph neural networks can be used to model the dependencies between features. Finally, feature selection operations are performed to retain feature information that is useful for the classification task and to remove redundant or noisy features. This can be achieved by applying methods such as L1 regularization, Dropout, or feature importance evaluation.

[0225] After the above processing, the second strategy labeling features corresponding to the two descriptors of basin area and average annual precipitation are obtained. These features contain deeper information than the first strategy labeling features and are used for subsequent classification tasks.

[0226] In step S203, the task of the computer system is to perform downsampling processing on the second strategy marking feature corresponding to each description item. The purpose of downsampling is to reduce the feature dimension, reduce the computational complexity and prevent overfitting.

[0227] Downsampling is usually achieved through pooling operations (such as maximum pooling, average pooling, etc.), dimensionality reduction techniques (such as principal component analysis PCA, linear discriminant analysis LDA, etc.), or feature selection methods (such as recursive feature elimination RFE, tree-based methods, etc.). These methods can map high-dimensional features to low-dimensional space while retaining as much useful information as possible. In practical applications, downsampling may be embedded in the model as part of the strategy labeling estimation algorithm, or it may be performed as a preprocessing step before model training.

[0228] In step S204, the task of the computer system is to classify the labeled sampling features using the strategy classifier of the strategy label prediction algorithm to obtain the predicted strategy label of the target reservoir. The strategy classifier is, for example, a neural network layer or a combination of a group of layers that can learn how to map input features (in this case, labeled sampling features) to output categories (i.e., predicted strategy labels). This process may involve linear classifiers (such as logistic regression, support vector machines, etc.), nonlinear classifiers (such as neural networks, decision trees, etc.), or ensemble learning methods (such as random forests, gradient boosting machines, etc.). In practical applications, the strategy classifier may be embedded in the model as the last part of the strategy label prediction algorithm. It receives the labeled sampling features obtained in the previous step as input and performs classification processing through its internal network structure.

[0229] In one implementation, step S201, extracting the strategy tag feature of each description item included in the strategy description data through a tag feature encoder of a strategy tag estimation algorithm, and obtaining a first strategy tag feature corresponding to each description item, including:

[0230] Step S2011: Determine the code corresponding to each description item included in the policy description data through the tag feature encoder of the policy tag estimation algorithm;

[0231] Step S2012: using a marking feature encoder to index the first strategy marking feature corresponding to each description item from a preset feature library according to the encoding symbol corresponding to each description item.

[0232] In step S2011, the computer system determines the coder corresponding to each descriptive item in the policy description data. The coder is a unique identifier used to map the descriptive item to the corresponding feature representation in the preset feature library. This step is the basis of the feature extraction process, which ensures that the descriptive item can be correctly indexed and converted. The determination of the coder usually depends on the type and content of the descriptive item. For numerical descriptive items, the coder may directly correspond to the binary representation or hash value of the numerical value; for textual descriptive items, the coder may be a unique identifier generated by a hash function or text embedding technology. In practical applications, the determination of the coder can take into account the semantic information and contextual relationship of the descriptive item. For example, for descriptive items with similar meanings but different expressions, text similarity algorithms (such as cosine similarity, Jaccard similarity, etc.) can be used to identify their similarities and assign them the same coder.

[0233] In step S2012, the computer system uses the coder determined in step S2011 to index the first strategy marking feature corresponding to each description item from the preset feature library. The preset feature library is a database containing a large number of pre-trained feature representations, which are learned from a large amount of historical data through a machine learning algorithm.

[0234] In practical applications, the preset feature library may contain multiple types of feature representations, such as numerical feature vectors, text embedding vectors, image feature maps, etc. These feature representations can be indexed and used according to different description item types. For example, continuing the above example, assume that a preset feature library containing numerical feature vectors and text embedding vectors has been constructed. Now, the coder determined in step S2011 will be used to index the first strategy marking feature corresponding to each description item. For the description item "watershed area", its coder is 1111101000. Find the numerical feature vector matching the coder in the preset feature library and use it as the first strategy marking feature corresponding to "watershed area". Assume that the feature vector found is [0.1, 0.3, 0.5, 0.2] (just an example). For the description item "average annual precipitation", its coder is 11001000. Similarly, find the numerical feature vector matching the coder in the preset feature library and use it as the first strategy marking feature corresponding to "average annual precipitation". Assume that the feature vector found is [0.2, 0.4, 0.1, 0.3] (just an example). For the description item "Policy Type", its code is abcdef1234567890. Search for the text embedding vector that matches the code in the preset feature library and use it as the first policy tag feature corresponding to "Policy Type". Assume that the text embedding vector found is [0.6, -0.1, 0.3, -0.2, 0.5] (the dimension may be adjusted according to the actual situation).

[0235] In another implementation, step S200 extracts strategy tag features from strategy description data by using a strategy tag prediction algorithm to obtain extracted strategy tag features, and performs strategy tag decoding on the extracted strategy tag features to obtain a predicted strategy tag of a target reservoir, including:

[0236] Step S210: extracting the strategy tag feature of each description item included in the strategy description data through the tag feature encoder of the strategy tag estimation algorithm, and obtaining the third strategy tag feature corresponding to each description item;

[0237] Step S220: embedding the third strategy tag feature corresponding to each description item through the embedder of the strategy tag estimation algorithm to obtain the fourth strategy tag feature corresponding to each description item;

[0238] Step S230: Decode the fourth strategy tag feature corresponding to each description item through the decoder of the strategy tag prediction algorithm to obtain the predicted strategy tag of the target reservoir.

[0239] In step S210, the task of the computer system is to use the tag feature encoder of the strategy tag estimation method to extract features from each descriptive item in the strategy description data. These descriptive items include, for example, basic basin characteristics (such as basin area, average annual precipitation, etc.), historical strategy tags (such as the first strategy tag, the second strategy tag, etc.) and other information related to the reservoir operation strategy.

[0240] The tag feature encoder is, for example, a neural network layer or a combination of layers that can convert the input description items into high-dimensional vector representations, namely the third strategy tag features. This conversion process usually involves nonlinear transformations and feature learning in order to capture the complex relationships and patterns in the description items. In practical applications, the tag feature encoder may adopt structures such as convolutional neural networks (CNN), recurrent neural networks (RNN), long short-term memory networks (LSTM), gated recurrent units (GRU) or Transformer. These structures can process different types of input data and extract useful feature information.

[0241] In step S210, the marker feature encoder receives this data item as input and processes it through its internal network structure. For example, a fully connected layer (FullyConnectedLayer) can be used as part of the marker feature encoder to convert each description item into a fixed-length vector representation. For the "watershed area" description item, assume that its output vector after processing by the fully connected layer is: [0.1, 0.2, 0.3, ..., 0.8]. This is a vector of length N (N is the preset vector length), which is regarded as the third strategy marker feature corresponding to "watershed area".

[0242] Similarly, for the description item of "average annual precipitation", the output vector after being processed by the fully connected layer may be: [0.4, 0.5, 0.6,..., 0.1]; this is also a vector of length N, which is regarded as the third policy marker feature corresponding to "average annual precipitation".

[0243] In step S220, the task of the computer system is to further process the third policy marker feature corresponding to each description item, that is, feature embedding. The purpose of feature embedding is to map the third policy marker feature into a lower-dimensional or higher-dimensional space to better capture the similarities and differences between features. The embedder is, for example, a neural network layer or a combination of a group of layers, which can perform non-linear transformation and dimensionality reduction / dimensionality increase processing on the input feature vector. In practical applications, the embedder may adopt techniques such as Autoencoder, Principal Component Analysis (PCA), t-SNE, etc. to implement feature embedding.

[0244] For example, continuing the above example, assume that the third policy marker features corresponding to the two description items of "basin area" and "average annual precipitation" have been obtained. Now, the embedder will be used to further process these features.

[0245] For example, an autoencoder can be used as part of the embedder. An autoencoder consists of an encoder and a decoder. The encoder compresses the input feature vector into a lower-dimensional latent space representation (i.e., the embedding vector), and the decoder then attempts to reconstruct the original input feature vector from the latent space representation. For the third policy marker feature corresponding to "basin area": [0.1, 0.2, 0.3,..., 0.8], after being processed by the encoder of the autoencoder, an embedding vector of length M (M < N) may be obtained as the fourth policy marker feature corresponding to "basin area": [0.5, 0.3,..., 0.2].

[0246] Similarly, for the third policy marker feature corresponding to "average annual precipitation", after being processed by the encoder of the autoencoder, an embedding vector of length M may also be obtained as the fourth policy marker feature corresponding to "average annual precipitation".

[0247] In step S230, the task of the computer system is to perform decoding processing on the fourth policy marker feature corresponding to each description item to generate the predicted policy marker of the target reservoir. The decoding process usually involves mapping the embedding vector back to the original policy marker space or a new policy marker space.

[0248] The decoder is, for example, a neural network layer or a combination of layers that can perform nonlinear transformation and classification processing on the embedded vector to generate the final prediction result. In practical applications, the decoder may be implemented using a fully connected layer, a softmax layer, or other classifiers.

[0249] For example, continuing the above example, assume that the fourth strategy tag features (i.e., embedding vectors) corresponding to the two descriptive items "watershed area" and "average annual precipitation" have been obtained. Now, the decoder will be used to decode these embedding vectors.

[0250] For example, you can use a fully connected layer followed by a softmax layer as part of the decoder. The fully connected layer maps the embedding vector into a higher-dimensional space, while the softmax layer transforms the vector in this space into a probability distribution representing the likelihood of different policy labels.

[0251] For the fourth strategy tag feature corresponding to "watershed area": ​​[0.5, 0.3, ..., 0.2], after being processed by the fully connected layer and the softmax layer, a probability distribution vector may be obtained, indicating the possibility of different strategy tags (such as "increase water storage", "reduce water release", etc.). For example: [0.1, 0.7, 0.2], this means that the "reduce water release" strategy has the highest probability (0.7), so it can be regarded as the prediction strategy tag of the target reservoir. Similarly, for the fourth strategy tag feature corresponding to "average annual precipitation", after being processed by the decoder, a probability distribution vector indicating the possibility of different strategy tags may be obtained, and the strategy tag with the highest probability is selected as the prediction result.

[0252] It should be noted that in the embodiment of the present invention, if the above-mentioned reservoir irrigation area water resources allocation 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 invention 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 invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), disk or optical disk, etc. Various media that can store program codes. In this way, the embodiment of the present invention is not limited to any specific hardware, software or firmware, or any combination of hardware, software, and firmware.

[0253] The above description is only an implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for allocating water resources in a reservoir irrigation area based on big data analysis, characterized in that: include: Obtain detailed information on the allocation of water resources from the target reservoir to the target irrigation area; Acquire the water resource allocation sequence characteristics of the target reservoir and the basic characteristics of the target watershed corresponding to the target reservoir from the allocation details information of the target reservoir; Using a target water resource allocation algorithm, combined with the water resource allocation sequence characteristics of the target reservoir and the basin basic characteristics of the target basin, the target irrigation area is estimated to obtain the demand area estimation information of the target irrigation area, and the demand area estimation information is an estimation result representing the target irrigation area including the irrigation demand area; The target water resource allocation algorithm is trained using the following steps: Acquire allocation detail information of a reservoir template and an irrigation demand training label of an irrigation district template corresponding to the allocation detail information, wherein the irrigation demand training label is a training label representing that the irrigation district template contains an irrigation demand area; Acquire the water resource allocation sequence characteristics of the reservoir template and the watershed basic characteristics of the watershed template corresponding to the reservoir template from the allocation detail information of the reservoir template; Using a basic water resource allocation algorithm, combined with the water resource allocation sequence characteristics and the basin basic characteristics, the irrigation demand of the irrigation area template is estimated to obtain the irrigation demand estimation information of the irrigation area template; the irrigation demand estimation information is an estimation result representing the irrigation demand area included in the irrigation area template; The basic water resource allocation algorithm is trained in combination with the irrigation demand training label and the irrigation demand estimation information to obtain a target water resource allocation algorithm; the degree of consistency between the irrigation demand optimization result of the irrigation area template output by the target water resource allocation algorithm and the irrigation demand training label is higher than the degree of consistency between the irrigation demand estimation information and the irrigation demand training label.

2. The method according to claim 1, characterized in that The water resource allocation sequence features include water resource allocation variables of the reservoir template at a first number of time points, a first number of irrigation nodes at the first number of time points, and irrigation district basic information of the irrigation district template; The basic water resource allocation algorithm is used to estimate the irrigation demand of the irrigation area template in combination with the water resource allocation sequence characteristics and the basin basic characteristics to obtain the irrigation demand estimation information of the irrigation area template, including: Averaging the water resource allocation variables at the first number of time points to obtain an average water resource allocation characteristic of the reservoir template; generating irrigation district description features describing the first number of irrigation nodes and the irrigation district template in combination with the first number of irrigation nodes and the irrigation district basic information of the irrigation district template; The basic water resource allocation algorithm is adopted, combined with the average water resource allocation characteristics, the irrigation area description characteristics, the water resource allocation sequence characteristics and the basin basic characteristics, to estimate the irrigation demand of the irrigation area template and obtain the irrigation demand estimation information of the irrigation area template.

3. The method according to claim 2, characterized in that The basic water resource allocation algorithm is used to estimate the irrigation demand of the irrigation area template in combination with the average water resource allocation characteristics, the irrigation area description characteristics, the water resource allocation sequence characteristics and the basin basic characteristics, to obtain the irrigation demand estimation information of the irrigation area template, including: Using the feature encoder of the basic water resource allocation algorithm, the water resource allocation sequence feature is encoded to obtain a water resource encoding vector for estimating the irrigation demand area; Using the integrator of the basic water resource allocation algorithm, the average water resource allocation feature and the basin basic feature are integrated to obtain an integrated feature; Using the feature changer of the basic water resources allocation algorithm, the integrated feature is subjected to feature change to obtain the main integrated feature in the integrated feature; The classifier of the basic water resource allocation algorithm is used, combined with the water resource coding vector, the main integrated features and the irrigation area description features, to estimate the irrigation demand of the irrigation area template, and obtain the irrigation demand estimation information of the irrigation area template.

4. The method according to claim 3, characterized in that The feature encoder of the basic water resource allocation algorithm is used to encode the water resource allocation sequence features to obtain a water resource coding vector for estimating the irrigation demand area, including: Using a feature encoder of the basic water resource allocation algorithm, generating a sequence encoding representation describing features of the water resource allocation sequence; Incorporating the distribution information of the first number of irrigation nodes into the sequence code representation to obtain a temporary code representation; Based on the first significance factor, the second significance factor and the third significance factor of the feature encoder, the temporary coding representation is respectively significance-weighted to obtain a first candidate coding representation, a second candidate coding representation and a third candidate coding representation; Determine a water resource code vector for estimating an irrigation demand area by combining the first candidate code representation, the second candidate code representation, and the third candidate code representation; The classifier using the basic water resource allocation algorithm, combined with the water resource coding vector, the main integrated features and the irrigation area description features, estimates the irrigation demand of the irrigation area template to obtain the irrigation demand estimation information of the irrigation area template, including: Using the classifier of the basic water resources allocation algorithm, the irrigation area description features are divided to obtain a second number of sub-irrigation area description features; One or more of the second number of sub-irrigation area description features include irrigation nodes, and do not include irrigation areas of the irrigation area template; Performing feature fusion on the water resource coding vector, the main integration feature, and the second number of sub-irrigation area description features to obtain a fusion feature; The classifier is used in combination with the fusion feature to estimate the irrigation demand of the irrigation area template, so as to obtain the irrigation demand estimation information of the irrigation area template.

5. The method according to claim 4, characterized in that The step of combining the first standby coding representation, the second standby coding representation, and the third standby coding representation to determine a water resource coding vector for estimating an irrigation demand area includes: multiplying the first candidate coded representation and the second candidate coded representation to obtain a significance vector describing that the irrigation area template includes an irrigation demand area; The product between the significance vector and the third candidate coded representation is combined to obtain a water resource coded vector for estimating the irrigation demand area.

6. The method according to claim 1, characterized in that The step of training the basic water resource allocation algorithm in combination with the irrigation demand training label and the irrigation demand estimation information to obtain a target water resource allocation algorithm includes: Determine the irrigation demand area estimation cost of the basic water resource allocation algorithm by combining the irrigation demand training label and the irrigation demand estimation information; Determining the stability of the basic water resource allocation algorithm in combination with the estimated cost of the irrigation demand area; When the basic water resource allocation algorithm is unstable, the basic water resource allocation algorithm is corrected in combination with the estimated cost of the irrigation demand area to obtain a target water resource allocation algorithm.

7. The method according to claim 6, characterized in that When the basic water resource allocation algorithm is unstable, the basic water resource allocation algorithm is corrected in combination with the estimated cost of the irrigation demand area to obtain a target water resource allocation algorithm, including: When the basic water resource allocation algorithm is unstable, the algorithm configuration parameters of the basic water resource allocation algorithm are optimized in combination with the estimated cost of the irrigation demand area to obtain an optimized basic water resource allocation algorithm; The optimized basic water resource allocation algorithm is used, combined with the water resource allocation sequence characteristics and the basin basic characteristics, to estimate the irrigation demand of the irrigation area template, and obtain the irrigation demand optimization result of the irrigation area template; the irrigation demand optimization result is the optimization result representing the irrigation demand area included in the irrigation area template; Determining whether the optimized basic water resource allocation algorithm is stable by combining the irrigation demand optimization result and the irrigation demand training label; When the optimized basic water resource allocation algorithm is stable, the optimized basic water resource allocation algorithm is used as the target water resource allocation algorithm.

8. The method according to claim 1, characterized in that The method further comprises: Acquire estimated information of a second number of demand areas corresponding to the target irrigation area, where the estimated information of the second number of demand areas is determined in combination with different allocation detail information corresponding to the target irrigation area; Determine, in combination with the estimated information of the second number of demand areas, a support coefficient describing the target irrigation area including the irrigation demand area; When the support coefficient is higher than the support coefficient, acquiring an irrigation network including the target irrigation area; In combination with the estimated information of the second number of demand areas, marking the irrigation demand areas of the target irrigation area in the irrigation network; The method further comprises: Acquire strategy reference data of the target reservoir, wherein the strategy reference data includes the basic characteristics of the watershed, a first strategy tag, and a second strategy tag, wherein the first strategy tag and the second strategy tag are determined according to the scheduling strategy of the target reservoir for irrigation demand in different monitoring cycles; Generate strategy description data corresponding to the target reservoir according to the basic characteristics of the watershed, the first strategy mark and the second strategy mark; The strategy mark feature is extracted from the strategy description data through a strategy mark estimation method to obtain the extracted strategy mark feature, and the extracted strategy mark feature is decoded to obtain the predicted strategy mark of the target reservoir.

9. The method according to claim 8, characterized in that The method of extracting strategy mark features from the strategy description data by using a strategy mark estimation algorithm to obtain extracted strategy mark features, and performing strategy mark decoding on the extracted strategy mark features to obtain a predicted strategy mark of the target reservoir includes: Extracting the strategy tag feature of each description item included in the strategy description data through a tag feature encoder of a strategy tag estimation algorithm to obtain a first strategy tag feature corresponding to each description item; Performing strategy tag feature learning on the first strategy tag feature corresponding to each description item by the learner of the strategy tag estimation algorithm to obtain the second strategy tag feature corresponding to each description item; Down-sampling the second strategy tag feature corresponding to each description item by the strategy tag estimation algorithm to obtain a tag sampling feature corresponding to the strategy description data; Classifying the label sampling features corresponding to the strategy description data by the strategy classifier of the strategy label prediction algorithm to obtain the prediction strategy label of the target reservoir; Alternatively, extracting strategy tag features from the strategy description data by using a strategy tag estimation method to obtain extracted strategy tag features, and performing strategy tag decoding on the extracted strategy tag features to obtain a predicted strategy tag of the target reservoir includes: Extracting the strategy tag feature of each description item included in the strategy description data through a tag feature encoder of the strategy tag estimation algorithm to obtain a third strategy tag feature corresponding to each description item; The embedder of the strategy tag estimation algorithm embeds the third strategy tag feature corresponding to each description item to obtain the fourth strategy tag feature corresponding to each description item; The fourth strategy mark feature corresponding to each description item is decoded by the decoder of the strategy mark prediction algorithm to obtain the predicted strategy mark of the target reservoir.

10. The method according to claim 9, characterized in that The tag feature encoder using the strategy tag estimation algorithm extracts the strategy tag feature from each description item included in the strategy description data to obtain a first strategy tag feature corresponding to each description item, including: Determine the code corresponding to each description item included in the policy description data by using the tag feature encoder of the policy tag estimation algorithm; The tag feature encoder is used to index the first strategy tag feature corresponding to each description item from a preset feature library according to the coders corresponding to each description item.

Images