A Method and System for Optimizing Operating Parameters and Data Governance in Water Conservancy Project Construction

Abstract

The present invention discloses a method and system for optimizing operation parameters and data governance in water conservancy project construction, including: First, through operations such as vector feature extraction and feature extraction on water volume scheduling parameters, multiple task types are determined and the confidence levels of their abnormal events are predicted respectively. After extracting the influence factor weighting coefficients and performing weighted processing, the overall abnormal event evaluation confidence level of the project is obtained. If there are abnormalities, real-time monitoring is carried out on each subsystem and key equipment, and preset governance strategies are pushed for optimization and adjustment to improve the operation and management efficiency of water conservancy projects.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular, to a method and system for optimizing operation parameters and data governance in water conservancy project construction. Background Art

[0002] In the operation and management of water conservancy projects, accurate water volume scheduling and reliable data governance are crucial. Traditional methods have problems such as insufficient analysis of water volume scheduling parameters and difficulty in comprehensively evaluating abnormal risks under various task types when dealing with complex working conditions and potential abnormal events. Summary of the Invention

[0003] The purpose of the present invention is to provide a method and system for optimizing operation parameters and data governance in water conservancy project construction.

[0004] In a first aspect, an embodiment of the present invention provides a method for optimizing operation parameters and data governance in water conservancy project construction, including:

[0005] Performing a vector feature extraction operation on the water volume scheduling parameters of the target water conservancy project to obtain a water volume scheduling vector representation corresponding to the water volume scheduling parameters;

[0006] Performing a feature extraction operation on the water volume scheduling vector representation to obtain a target water volume scheduling feature corresponding to the water volume scheduling vector representation;

[0007] Determining multiple task types corresponding to the water volume scheduling parameters, and respectively performing abnormal event prediction processing on the target water volume scheduling feature according to each task type to obtain an abnormal event prediction confidence corresponding to each task type;

[0008] Performing an influencing factor extraction process on the water volume scheduling vector representation to obtain an influencing factor weighting coefficient corresponding to each task type;

[0009] Weighting the abnormal event prediction confidence corresponding to each task type based on the influencing factor weighting coefficient corresponding to each task type to obtain a weighted abnormal event prediction coefficient corresponding to each task type; an influencing factor conversion value corresponding to a task type is used to weight the abnormal event prediction confidence corresponding to the corresponding task type;

[0010] Determining an abnormal event evaluation confidence of the target water conservancy project according to the weighted abnormal event prediction coefficient corresponding to each task type;

[0011] In the case where the abnormal event evaluation confidence indicates that the target water conservancy project has an abnormal event, collecting real-time monitoring data of each subsystem and key equipment of the target water conservancy project, and combining the abnormal event evaluation confidence to push a preset governance strategy to each subsystem for optimization and adjustment.

[0012] In a second aspect, an embodiment of the present invention provides a server system, including a server for executing the method described in the first aspect.

[0013] Compared with the prior art, the beneficial effects provided by the present invention include: By using a method and system for optimizing operation parameters and data governance in water conservancy project construction disclosed by the present invention, through operations such as vector feature extraction and feature extraction of water volume scheduling parameters, multiple task types are determined and the confidence levels of their abnormal events are predicted respectively. After extracting the influence factor weighting coefficients and performing weighted processing, the overall abnormal event assessment confidence level of the project is obtained. If there is an abnormality, real-time monitoring is carried out on each subsystem and key equipment, and a preset governance strategy is pushed for optimization and adjustment, thereby improving the operation management efficiency of water conservancy projects. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the embodiments. It should be understood that the following drawings only show some embodiments of the present invention, and thus should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can also be obtained based on these drawings without creative efforts.

[0015] Figure 1 It is a schematic flow chart of the steps of the method for optimizing operation parameters and data governance in water conservancy project construction provided by the embodiment of the present invention;

[0016] Figure 2 It is a schematic block diagram of the structure of the computer device provided by the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Usually, the components of the embodiments of the present invention described and illustrated in the drawings here can be arranged and designed in various different configurations.

[0018] The following will detail the specific implementation manners of the present invention with reference to the drawings.

[0019] To solve the technical problems in the foregoing background art, Figure 1 It is a schematic flow chart of the method for optimizing operation parameters and data governance in water conservancy project construction provided by the embodiment of the present disclosure. The following will introduce this method for optimizing operation parameters and data governance in water conservancy project construction in detail.

[0020] Step S201: Perform a vector feature extraction operation on the water volume scheduling parameters of the target water conservancy project to obtain the corresponding water volume scheduling vector representation of the water volume scheduling parameters;

[0021] Step S202: Perform a feature extraction operation on the water volume scheduling vector representation to obtain the corresponding target water volume scheduling features of the water volume scheduling vector representation;

[0022] Step S203: Determine multiple task types corresponding to the water volume scheduling parameters, and perform abnormal event prediction processing on the target water volume scheduling features respectively according to each task type to obtain the abnormal event prediction confidence levels corresponding to each task type;

[0023] Step S204: Perform an influence factor extraction process on the water volume scheduling vector representation to obtain the influence factor weighting coefficients corresponding to each task type;

[0024] Step S205: Weight the abnormal event prediction confidence levels corresponding to each task type based on the influence factor weighting coefficients corresponding to each task type to obtain the weighted abnormal event prediction coefficients corresponding to each task type; the influence factor conversion value corresponding to a task type is used to weight the abnormal event prediction confidence level corresponding to the corresponding task type;

[0025] Step S206: Determine the abnormal event evaluation confidence level of the target water conservancy project according to the weighted abnormal event prediction coefficients corresponding to each task type;

[0026] Step S207: When the abnormal event evaluation confidence level indicates that there is an abnormal event in the target water conservancy project, collect real-time monitoring data of each subsystem and key equipment of the target water conservancy project, and combine the abnormal event evaluation confidence level to push a preset governance strategy to each subsystem for optimization and adjustment.

[0027] In an embodiment of the present invention, exemplarily, the server first starts to collect data related to water volume scheduling parameters from various data sources of the target water conservancy project. The water conservancy project includes multiple reservoirs, multiple river connection segments, and numerous water conservancy facilities, and the data sources are rich and diverse. Real-time water level data can be obtained from the water level monitoring sensors of each reservoir, which is updated once an hour, and details the water level heights at different times, such as the water level at 8 am is 100 meters, and at 9 am is 100.2 meters during a specific period. The river flow monitoring stations can provide river flow data, such as the flow rate of a certain main river connection segment is 500 cubic meters per second at a certain moment. At the same time, data such as the opening degree of the gate and the operating power of the pumping station can also be obtained from the operation control system of the water conservancy facilities. For example, the gate opening degree is in a 30% open state during a certain period, and the operating power of the pumping station is 200 kilowatts. Since these data types, formats, and dimensions are different, the server needs to organize and preprocess them to unify the format for subsequent operations. For example, the water level data is unified in meters, and the flow rate data is standardized to the format of cubic meters per second. Then, a suitable feature encoding method is adopted, such as one-hot encoding for the opening degree of the gate in different states, and numerical encoding for data such as water level, flow rate, and operating power of the pumping station. The preprocessed water volume scheduling parameters are converted into vector form, and a water volume scheduling vector representation such as [100 (numerical encoding of water level), 500 (numerical encoding of flow rate), 0.3 (value after one-hot encoding conversion of gate opening degree), 200 (numerical encoding of operating power of the pumping station)] is obtained, completing the conversion from the original parameters to the vector representation, laying a foundation for subsequent analysis. For the water level data, taking the past 24 hours as an example, if the water level shows a trend of rising slowly first and then falling slowly, by calculating the water level difference between adjacent time points and analyzing its change pattern, target water volume scheduling features reflecting the water level change trend can be obtained, such as the average water level rising rate is 0.1 meters per hour in a few hours, and the average falling rate is 0.08 meters per hour in the subsequent hours. For the flow rate data, its seasonal and periodic characteristics are analyzed. The area where the water conservancy project is located is a rainy season in summer. After analyzing the flow rate data over the years, it is found that the flow rate will rise periodically from June to August every year, with an average increase of about 30% compared to other months. This is the target water volume scheduling feature corresponding to the flow rate data. Relevant features can also be extracted from data such as the opening degree of the gate and the operating power of the pumping station, such as the frequent adjustment frequency feature of the gate opening degree and the stable operation duration feature of the operating power of the pumping station. According to the actual operation requirements and functions of the water conservancy project, multiple task types such as flood control, irrigation water supply, power generation, and ecological water replenishment are determined. For the flood control task, using the extracted target water volume scheduling features as input, a decision tree algorithm is used to construct an abnormal event prediction model. The input features include the water level change trend feature (too fast water level rising rate may indicate an increased flood risk) and the seasonal and periodic characteristics of the river flow rate (abnormally high flow rate in the non-rainy season may be a flood hazard). After training the model with historical flood control data, the probability of abnormal events under the flood control task can be predicted according to the current input features.Similarly, for the irrigation water supply task, a corresponding model is constructed using the support vector machine algorithm. The input features include water level data (low water levels may lead to insufficient water supply) and the frequency feature of frequent adjustment of the gate opening. After inputting the current features, the prediction confidence of abnormal events of insufficient water supply can be predicted. For other task types such as power generation tasks and ecological water replenishment tasks, corresponding abnormal event prediction models are also constructed respectively, and the corresponding prediction confidence of abnormal events is calculated based on the target water volume scheduling features input for each task, so as to comprehensively evaluate the risk of abnormal events that may occur under each task type. The server further analyzes the water volume scheduling vector representation and extracts the corresponding influence factor weighting coefficients for each task type. In the flood control task, the water level change trend feature is extremely critical for flood control, and the influence factor weighting coefficient may be 0.5; the seasonal feature of river flow has a slightly lower degree of influence, and the coefficient may be 0.3; the frequency feature of frequent adjustment of the gate opening has a relatively small influence, and the coefficient is 0.1; the influence of the pump station operation power feature is not significant, and the coefficient is 0.1. In the irrigation water supply task, the influence factor weighting coefficient of water level data may be relatively high, such as 0.4; the frequency feature of frequent adjustment of the gate opening is more important, and the coefficient is 0.3; the seasonal feature of flow data has a relatively small influence, and the coefficient is 0.2; the influence of the pump station operation power feature is not significant, and the coefficient is 0.1. For other task types, the influence factor weighting coefficients are also determined according to the importance degree of each feature for the corresponding task. Then, according to the weighting formula, the influence factor weighting coefficients corresponding to each feature are weighted and calculated with the prediction confidence of abnormal events to obtain the corresponding weighted abnormal event prediction coefficients for each task type. Finally, to determine the evaluation confidence of abnormal events for the target water conservancy project, methods such as weighted average or simple average can be used. For example, if the calculated evaluation confidence is 0.6, it indicates a relatively high possibility of abnormal events. For the reservoir subsystem, increase the data collection frequency of reservoir water level, water temperature, water quality, etc., from once per hour to once every half hour. If the water level continues to rise and the evaluation confidence of abnormal events is high, push the preset treatment strategy, such as increasing the gate opening to lower the reservoir water level, and closely monitor the water level change. For the irrigation water supply subsystem, strengthen the data collection of irrigation channel flow, water quality, and the operation status of irrigation equipment. If the risk of insufficient water supply is high, push strategies such as giving priority to ensuring water supply in important crop planting areas, reasonably adjusting the water supply time and flow distribution, and checking for equipment failures. For the power generation subsystem, increase the data collection intensity of generator set power, temperature, vibration, etc. If the power generation efficiency may be affected, the pushed strategy is to comprehensively check the generator set, adjust the power generation load, and ensure the best operating state. For the ecological water replenishment subsystem, strengthen the data collection of the flow, water quality, and ecological environment indicators of the replenishment river. If there may be problems with ecological water replenishment, the pushed strategies include adjusting the water replenishment time and flow, and strengthening the monitoring of the surrounding ecological environment. Through such comprehensive real-time monitoring data collection and strategy pushing for optimization and adjustment, the server can effectively ensure the safe, stable, and efficient operation of the target water conservancy project and calmly handle various possible abnormal event situations.In summary, through the systematic processing, analysis of water conservancy project data, and the subsequent monitoring and strategy adjustment based on this, the accurate control and effective guarantee of the operation status of water conservancy projects can be achieved.

[0028] In the embodiment of the present invention, the operation of extracting the vector feature of the water volume scheduling parameter of the target water conservancy project to obtain the corresponding water volume scheduling vector representation of the water volume scheduling parameter can be implemented through the following examples.

[0029] Perform feature mapping processing on the water volume scheduling parameter to obtain the mapped feature parameter of the water volume scheduling parameter;

[0030] Extract the water volume scheduling vector representation from the mapped feature parameter to obtain the water volume scheduling vector representation of the mapped feature parameter.

[0031] In the embodiment of the present invention, for example, the server first receives the original data from each monitoring point and device, such as the water level values at different times in the reservoir, the flow data at various positions in the river, the gate opening information, and the pump station operation power data, etc. These data have different formats and dimensions and are difficult to directly use for in-depth analysis. Therefore, the server conducts feature mapping processing. For the water level data, a mapping interval is set (such as [0, 10]), and through a linear mapping function, it is converted into comparable mapped feature parameters. For example, a water level of 120 meters can be mapped to 6. Since the river flow data has a large variation range, a logarithmic mapping function is constructed based on the historical flow maximum and minimum values for processing. For example, a flow rate of 800 cubic meters per second can be mapped to 4.5 after mapping. The gate opening information is processed through normalization, and an opening of 40% is mapped to 0.4. The pump station operation power data is linearly mapped according to its historical power range, and a power of 300 kilowatts can be mapped to 0.5. Through these processes, the original diverse water volume scheduling parameters are converted into unified and comparable mapped feature parameters. After completing the feature mapping processing, the server performs the work of extracting the water volume scheduling vector representation. Taking this water conservancy project system as an example, the mapped feature parameters of each parameter have been obtained, such as a water level of 6, a flow rate of 4.5, a gate opening of 0.4, and a pump station operation power of 0.5. First, determine the dimension of the vector representation. If the above four factors are mainly considered, the dimension is 4. Then, place each mapped feature parameter into the vector in order to form a water volume scheduling vector representation in the form of [6, 4.5, 0.4, 0.5]. In practice, more factors may be considered, such as the precipitation in meteorological data. Perform feature mapping processing on the precipitation data. Assuming the mapped parameter is 3, the dimension of the extended vector is 5, and the water volume scheduling vector representation becomes [6, 4.5, 0.4, 0.5, 3]. If other factors such as water quality conditions are considered, after similar processing, their mapped feature parameters are placed into the vector to further improve the vector representation. Through the above operations, the server can accurately extract the vector representation reflecting the water volume scheduling situation of the target water conservancy project, providing convenience for subsequent analysis and processing and making it more efficient.

[0032] In the embodiment of the present invention, the feature mapping process for the water volume scheduling parameter to obtain the mapped feature parameter of the water volume scheduling parameter can be implemented through the following examples.

[0033] When the water volume scheduling parameter of the target water conservancy project is obtained, obtain the target abnormal event evaluation model for evaluating the abnormal events of the target water conservancy project; the target abnormal event evaluation model includes a first abnormal event evaluation sub-model;

[0034] Based on the first abnormal event evaluation sub-model, determine the demand information corresponding to the water volume scheduling parameter, and extract the target water use demand parameter of the water volume scheduling parameter according to the demand information;

[0035] Based on the first abnormal event evaluation sub-model, perform a mapping process on the target water use demand parameter to obtain the mapped process data corresponding to the target water use demand parameter;

[0036] Based on the first abnormal event evaluation sub-model, perform a standardization operation on the mapped process data corresponding to the target water use demand parameter to obtain the mapped feature parameter of the water volume scheduling parameter.

[0037] In an embodiment of the present invention, by way of example, the target water conservancy project is a large-scale comprehensive hub, covering facilities such as multiple reservoirs and complex irrigation channels, and serving many water use requirements. As the core of data processing, the server continuously collects various water volume scheduling parameters, such as real-time water level data updated every 15 minutes for each reservoir, irrigation channel flow data (fluctuating with seasons and demands), sluice opening information, and pump station operation power data, etc. At the same time, the server stores a target abnormal event evaluation model constructed over a long period, in which the first abnormal event evaluation sub-model focuses on the evaluation of abnormalities related to water use requirements. After the server obtains the relevant parameters and the model, it processes them according to the first abnormal event evaluation sub-model. Taking agricultural irrigation as an example, the model first analyzes the water volume scheduling parameters to determine the demand information, and clarifies the approximate irrigation water demand of the main crops at the current growth stage. Then, it accurately extracts the target water use demand parameters, such as screening corresponding data from the irrigation channel flow data, judging the water level height value that the reservoir needs to maintain according to the irrigation demand, and extracting the sluice opening value that affects the distribution and regulation of irrigation water. After extracting the parameters, mapping processing is performed. For the irrigation channel flow parameters, a mapping function is constructed according to the historical flow distribution law; the reservoir water level parameters are mapped according to the designed water level range and historical fluctuations; the sluice opening parameters are normalized. After obtaining the respective data through the mapping processing, a standardization operation is then performed. Taking each parameter in the agricultural irrigation scenario as an example, according to the standardization rule of unifying the mapped data into a normal distribution interval with a mean of 0 and a standard deviation of 1, and based on the overall mean and standard deviation of each parameter, through the standardization formula calculation, the mapped characteristic parameters of the final water volume scheduling parameters such as 0.67 (irrigation channel flow), 0.5 (reservoir water level), 0.5 (sluice opening) are obtained. These parameters are in a unified standard form to ensure the comparability and accuracy of subsequent analysis and processing.

[0038] In an embodiment of the present invention, the first abnormal event evaluation sub-model includes a data access layer; the target water use demand parameters include multiple target water use demand parameters, and among the multiple target water use demand parameters, there is a target key water use demand parameter;

[0039] Performing mapping processing on the target water use demand parameters based on the first abnormal event evaluation sub-model to obtain the corresponding mapped processing data of the target water use demand parameters can be implemented through the following examples.

[0040] According to the demand information corresponding to the target key water use demand parameter, determine the transform domain corresponding to the target key water use demand parameter; the transform domain includes multiple conversion values, and each conversion value has a corresponding target water use demand numerical range;

[0041] Based on the data access layer, determine the target water demand numerical range corresponding to the target key water demand parameter, and use the conversion value corresponding to the target water demand numerical range of the target key water demand parameter as the mapping processing data corresponding to the target key water demand parameter.

[0042] In an embodiment of the present invention, exemplarily, during daily operation, the server collects a large amount of water volume scheduling parameter data from various monitoring devices and systems. Among them, the target water demand parameters include multiple aspects, such as agricultural irrigation water demand parameters, urban residential water demand parameters, industrial water demand parameters, etc. Here, we focus on the target key water demand parameters, assuming they are the irrigation water demand parameters for a large-scale crop planting area. The server, through information docking with the agricultural department and analysis of the historical water use data of this planting area, etc., clarifies the demand information corresponding to the target key water demand parameters. For example, this planting area grows rice, and it is currently in a critical growth period. According to the growth characteristics and planting area of rice, it is understood that a stable and sufficient amount of irrigation water is required every day at this stage, and the quality of the irrigation water needs to meet certain standards. At the same time, the irrigation volume needs to be flexibly adjusted according to weather conditions (such as whether there is rainfall recently). Based on the above demand information, the server uses the relevant rules and algorithms in the first abnormal event evaluation sub-model to determine the transform domain corresponding to the target key water demand parameters. For the irrigation water demand parameters of this rice planting area, the determination of the transform domain considers multiple factors. For example, from the perspective of water volume, according to historical irrigation data and the water demand law of different growth stages of rice, multiple conversion values and their respective corresponding target water demand numerical ranges are determined. Suppose the target water demand numerical range corresponding to one of the conversion values is that the daily irrigation water volume is between 5000 cubic meters and 8000 cubic meters. This means that when the irrigation water volume is in this interval, it corresponds to a water use situation state represented by this conversion value; the numerical range corresponding to another conversion value can be that the daily irrigation water volume is between 8001 cubic meters and 10000 cubic meters, representing another different water use situation, such as the situation where the irrigation volume needs to be increased during a relatively dry period. From the perspective of water quality, there are also corresponding conversion values and numerical range settings. For example, when the pH value of the water quality is between 6.5 and 7.5, and the oxygen content is between 5 mg / L and 8 mg / L, it corresponds to a specific conversion value, representing a state where it meets the normal growth of rice and the quality of the irrigation water meets the standard. The data access layer of the server continuously receives real-time data on the irrigation water of this planting area from various monitoring points. For example, the flow data of the current irrigation channel is obtained from the flow monitor of the irrigation channel, and from this, the daily actual water volume irrigated to this planting area can be calculated; the index data such as the pH value and oxygen content of the water quality is obtained from the water quality monitoring equipment. Suppose the daily actual irrigation water volume of this planting area obtained through the data access layer is 6500 cubic meters, the pH value of the water quality is 7.0, and the oxygen content is 6 mg / L.According to the previously determined transform domain, the server finds that the irrigation water volume of 6500 cubic meters is within the target water consumption demand value range of "the daily irrigation water volume is between 5000 cubic meters and 8000 cubic meters", and the corresponding conversion value is A; the water quality situation also conforms to the numerical range corresponding to a certain previously set conversion value, and its corresponding conversion value is B. Then, the server uses the conversion values (here A and B) corresponding to the target water consumption demand value range corresponding to the target key water consumption demand parameter as the mapping processing data corresponding to the target key water consumption demand parameter. These mapping processing data will be further used for subsequent analysis and processing to evaluate whether there are abnormal events and make corresponding scheduling decisions, etc.

[0043] In the embodiment of the present invention, the target water volume scheduling feature is determined by the feature encoder in the first abnormal event evaluation sub-model performing a feature extraction operation on the water volume scheduling vector representation; the first abnormal event evaluation sub-model is the target abnormal event evaluation model for evaluating abnormal events of the target water conservancy project; the embodiment of the present invention also provides the following implementation manners.

[0044] Obtain a first sample training array; the first training instance included in the first sample training array is a training instance without an abnormal event target value;

[0045] Obtain multiple training targets for the first original abnormal event evaluation sub-model;

[0046] Based on the first original abnormal event evaluation sub-model to process the first training instance, obtain the sample water volume scheduling features corresponding to each training target, and determine the multiple error parameters corresponding to the multiple training targets according to the sample water volume scheduling features; one training target corresponds to one error parameter;

[0047] Determine a first target error parameter according to the multiple error parameters, optimize the model parameters of the first original abnormal event evaluation sub-model according to the first target error parameter, and use the optimized first original abnormal event evaluation sub-model as the first abnormal event evaluation sub-model.

[0048] In an embodiment of the present invention, exemplarily, the server uses the feature encoder in the first abnormal event evaluation sub-model to perform feature extraction operations on the obtained water volume scheduling vector representation. For vector values in the vector representation that involve information such as reservoir water level, river flow, and gate opening, the feature encoder will deeply analyze and extract target water volume scheduling features such as water level change trend, seasonal fluctuation characteristics of flow, and gate opening adjustment frequency. Then, the server carefully selects and organizes the first sample training array from the historical operation data of water conservancy projects. The first training instances in it are not configured with abnormal event target values, but are only data records related to water volume scheduling at different time periods, covering actual situations such as the water level, flow, and gate opening at that time. At the same time, for the first original abnormal event evaluation sub-model, multiple training objectives are determined, such as accurately predicting flood inundation in flood control scenarios, insufficient water supply in irrigation scenarios, and a significant decline in power generation efficiency in power generation scenarios. Subsequently, the first training instances in the first sample training array are sequentially input into the first original abnormal event evaluation sub-model for processing. Taking an instance as an example, the actual situation corresponding to its water volume scheduling vector representation is that the reservoir water level rises slowly, the river flow fluctuates normally, and the gate opening has been adjusted less recently. After model processing, corresponding sample water volume scheduling features are obtained for each training objective. For example, features such as the water level rise rate are extracted under the flood control training objective. Then, by comparing these sample water volume scheduling features with the ideal situations of the corresponding training objectives, multiple error parameters are determined. Each training objective has a corresponding error parameter. For example, under the flood control training objective, a deviation error parameter is generated when the actual water level rise rate is higher than the predicted normal rise rate. Finally, based on the determined multiple error parameters, the first target error parameter is calculated through a specific algorithm, comprehensively reflecting the overall error situation of the model. Then, the parameters of the first original abnormal event evaluation sub-model are adjusted and optimized using an optimization algorithm. After multiple rounds of optimization to reduce the error, the optimized model is used as the first abnormal event evaluation sub-model to enable it to more accurately evaluate abnormal events of target water conservancy projects.

[0049] In an embodiment of the present invention, the multiple training objectives include a random deletion training objective;

[0050] Processing the first training instance based on the first original abnormal event evaluation sub-model to obtain sample water volume scheduling features corresponding to each training objective, and determining multiple error parameters corresponding to the multiple training objectives according to the sample water volume scheduling features, including:

[0051] Performing feature mapping processing on the first training instance based on the first original abnormal event evaluation sub-model to obtain the first mapping feature parameter instance of the first training instance; the first mapping feature parameter instance includes mapping processing data instances on multiple index marks;

[0052] Randomly select a first index tag to be randomly deleted from the multiple index tags corresponding to the first mapping feature parameter instance;

[0053] Perform a random deletion process on the mapping processing data instance on the first index tag in the first mapping feature parameter instance according to the random deletion information to obtain a first feature parameter instance;

[0054] Based on the first original abnormal event evaluation sub-model, extract the water volume scheduling vector representation of the first feature parameter instance to obtain the first water volume scheduling vector representation instance of the first feature parameter instance, and perform a feature extraction operation on the first water volume scheduling vector representation instance to obtain the first sample water volume scheduling feature corresponding to the first water volume scheduling vector representation instance; the first sample water volume scheduling feature belongs to the sample water volume scheduling feature;

[0055] Based on the execution unit corresponding to the random deletion training target, perform a random deletion inference process on the first sample water volume scheduling feature to obtain the inference mapping processing data corresponding to the random deletion information in the first feature parameter instance;

[0056] Determine the error parameter corresponding to the random deletion training target according to the mapping processing data instance on the first index tag in the first mapping feature parameter instance and the inference mapping processing data.

[0057] In an embodiment of the present invention, exemplarily, the server selects a first training instance from the historical operation data of the water conservancy project. This first training instance contains rich information related to water volume scheduling, such as the water level values of each reservoir at a specific moment, the flow data of each irrigation channel, the opening conditions of different sluice gates, and the operating power of the pumping stations, etc. The server inputs this first training instance into the first original abnormal event evaluation sub-model, and the model first performs feature mapping processing on it. Taking the reservoir water level data as an example, assume that the water level of a certain reservoir at this moment is 120 meters. The model maps the 120-meter water level to a value within a specific interval (such as [0,1]) through a specific mapping function (such as a linear mapping function) according to the range of historical water level data (such as the lowest water level of 80 meters and the highest water level of 150 meters). Assume the mapped value is 0.6. Similarly, for the flow data of the irrigation channel, the opening conditions of the sluice gate, and the operating power of the pumping station, etc., the model also uses corresponding mapping methods to process them respectively, and finally obtains the first mapped feature parameter instance of the first training instance. This first mapped feature parameter instance contains mapped processing data instances on multiple index tags. For example, the mapped processing data instance with the index tag "water level" is 0.6, and the mapped processing data instance with the index tag "flow" is another mapped value, and so on. After obtaining the first mapped feature parameter instance, the server randomly selects a first index tag to be randomly deleted from the multiple index tags corresponding to this instance. Assume the index tags of the first mapped feature parameter instance are "water level", "flow", "sluice gate opening", "pumping station operating power", etc. The server uses tools such as a random number generator to randomly select the index tag "sluice gate opening" as the first index tag. After determining that the first index tag is "sluice gate opening", the server performs random deletion processing on the mapped processing data instance of the index tag "sluice gate opening" in the first mapped feature parameter instance according to the random deletion information. For example, the original mapped processing data instance of "sluice gate opening" is 0.4 (assume the value obtained after the previous mapping processing), and now according to the rules of random deletion processing, this value is deleted from the first mapped feature parameter instance to obtain the first feature parameter instance. At this time, the first feature parameter instance no longer contains the mapped processing data instance corresponding to the index tag "sluice gate opening". The server then inputs the first feature parameter instance into the first original abnormal event evaluation sub-model again. The model first extracts the water volume scheduling vector representation of the first feature parameter instance, and combines the mapped processing data instances corresponding to the remaining index tags into a vector in a certain order to obtain the first water volume scheduling vector representation instance of the first feature parameter instance. For example, if the remaining index tags are "water level" and "flow", and their corresponding mapped processing data instances are 0.6 and another value respectively, then the first water volume scheduling vector representation instance can be [0.6, that value].Then, the model performs a feature extraction operation on the first water volume scheduling vector representation instance, and extracts features from this vector representation instance that can reflect the characteristics and laws of water volume scheduling. For example, the water level change trend features (such as the rising rate, falling rate, etc.) are extracted from the numerical value corresponding to the water level, and the seasonal features, periodic features, etc. of the flow rate are extracted from the numerical value corresponding to the flow rate, to obtain the corresponding first sample water volume scheduling features of the first water volume scheduling vector representation instance. There is an execution unit in the server corresponding to the random deletion training target. This execution unit will receive the first sample water volume scheduling features and perform random deletion inference processing on them according to the preset rules and algorithms. Suppose the first sample water volume scheduling features include the water level change trend features (such as the rising rate is 0.1 m / h) and the seasonal features of the flow rate (such as the flow rate in summer is 30% higher than that in winter), etc. The execution unit may perform random deletion inference on a certain feature according to the requirements of the random deletion training target. For example, with a certain probability (assuming this probability is set according to historical data and model training requirements), the execution unit decides to perform random deletion inference processing on the seasonal features of the flow rate, and through a specific inference algorithm (determined according to the structure and training method of the model), obtains the inference mapping processing data corresponding to the random deletion information in the first feature parameter instance. Suppose after the inference processing, the obtained inference mapping processing data represents that there is a new inference value for the change of the flow rate in a certain season (this value is obtained based on the inference and data processing of the model). Finally, the server determines the error parameter corresponding to the random deletion training target according to the mapping processing data instance on the first index tag in the first mapping feature parameter instance (which was deleted earlier, here refers to the original value, that is, the mapping processing data instance of "gate opening" 0.4) and the inference mapping processing data (such as the inference value for the change of the flow rate in a certain season obtained earlier). Specifically, the server will compare the actual situation represented by the original mapping processing data instance of "gate opening" with the inference situation represented by the inference mapping processing data obtained after the random deletion inference processing. If there is a large difference between the two, for example, the original gate opening has a certain impact on the water volume scheduling, and after the inference, the situation of the flow rate changes greatly, resulting in the overall evaluation of the water volume scheduling not matching the actual situation, then a large error parameter will be generated. This error parameter reflects the deviation degree between the result of the model's processing of the first training instance and the actual situation under the random deletion training target, so that the first original abnormal event evaluation sub-model can be further optimized according to this error parameter, making it more accurate in evaluation and prediction when facing similar situations.

[0058] In the embodiment of the present invention, the multiple training targets include a local replacement training target;

[0059] Processing the first training instance based on the first original anomaly event evaluation sub-model to obtain sample water volume scheduling features corresponding to each training target, and determining a plurality of error parameters corresponding to the plurality of training targets according to the sample water volume scheduling features can be implemented through the following examples.

[0060] Performing feature mapping processing on the first training instance based on the first original anomaly event evaluation sub-model to obtain a first mapped feature parameter instance of the first training instance; the first mapped feature parameter instance includes mapped processing data instances on a plurality of index tags;

[0061] Randomly selecting a second index tag to be locally permuted from the plurality of index tags corresponding to the first mapped feature parameter instance;

[0062] Performing local permutation processing on the mapped processing data instance on the second index tag in the first mapped feature parameter instance according to preset element information to obtain a second feature parameter instance; the preset element information is different from the mapped processing data instance on the second index tag;

[0063] Extracting a water volume scheduling vector representation of the second feature parameter instance based on the first original anomaly event evaluation sub-model to obtain a second water volume scheduling vector representation instance of the second feature parameter instance, and performing a feature extraction operation on the second water volume scheduling vector representation instance to obtain a second sample water volume scheduling feature corresponding to the second water volume scheduling vector representation instance; the second sample water volume scheduling feature belongs to the sample water volume scheduling features;

[0064] Performing local permutation inference processing on the second sample water volume scheduling feature based on an execution unit corresponding to the local permutation training target to obtain local permutation inference results corresponding to the plurality of index tags;

[0065] Determining an error parameter corresponding to the local permutation training target according to the second index tag and the local permutation inference results corresponding to the plurality of index tags.

[0066] In an embodiment of the present invention, exemplarily, the server selects a first training instance from historical operation data, which covers water volume scheduling data such as reservoir water level, river flow rate, sluice opening degree, and pumping station operation power. For example, at a certain moment, the reservoir water level is 130 meters, the river flow rate is 800 cubic meters per second, the sluice opening degree is 30%, and the pumping station operation power is 250 kilowatts. After this instance is input into the first original abnormal event evaluation sub-model, feature mapping processing is performed. Taking the reservoir water level as an example, it is linearly mapped to the [0, 1] interval according to the historical water level range (80 meters to 150 meters), and the water level of 130 meters can obtain a mapped processing data instance of 0.6 after mapping. Data such as river flow rate is also processed according to its respective mapping rules to obtain a first mapped feature parameter instance, including mapped processing data instances corresponding to multiple index tags such as "water level" and "flow rate". Then, the server randomly selects a second index tag to be locally replaced from these index tags, such as selecting "flow rate". According to the preset element information, local replacement processing is performed on the mapped processing data instance of the "flow rate" index tag. The originally mapped value corresponding to a flow rate of 800 cubic meters per second (assumed to be 0.4) is replaced with a mapped processing data instance (assumed to be 0.3) obtained according to the historical typical flow rate value (such as 500 cubic meters per second) through the same mapping rule, so as to obtain a second feature parameter instance. Subsequently, the second feature parameter instance is input into the model again. First, water volume scheduling vector representation extraction is performed, and the mapped processing data instances corresponding to each index tag are combined in sequence into a vector to obtain a second water volume scheduling vector representation instance, such as [0.6, 0.3, 0.3, 0.25]. Then, a feature extraction operation is performed to extract features such as the water level change trend, the features related to the replaced flow rate, the influence features of the sluice opening degree change, and the stability features of the pumping station operation power, to obtain a second sample water volume scheduling feature. The execution unit corresponding to the local replacement training target in the server receives the second sample water volume scheduling feature and performs local replacement inference processing. For the features related to each index tag, such as inferring the change trend of "water level" under the replaced flow rate, and inferring the subsequent change situation of "flow rate", etc., multiple local replacement inference results are obtained. Finally, an error parameter is determined according to the second index tag "flow rate" and the local replacement inference results corresponding to each index tag. When not replaced, the actual operation data has its own rules, and if there are situations such as an obvious upward trend in the water level, and significant differences in the adjustment of the sluice opening degree and pumping station operation power compared with before replacement after replacement, the degree of difference between the actual situation and the result after replacement inference processing is quantified as an error parameter, which is used to reflect the deviation between the processing result of the model for the first training instance and the actual situation, so as to optimize the first original abnormal event evaluation sub-model subsequently. Through such a data processing and analysis process, the model is continuously improved to enable it to more accurately serve the operation evaluation and management of water conservancy projects.

[0067] In an embodiment of the present invention, among the multiple training targets, there is an interference addition training target;

[0068] Processing the first training instance based on the first original abnormal event evaluation sub-model to obtain the sample water volume scheduling features corresponding to each training objective, and determining a plurality of error parameters corresponding to the plurality of training objectives according to the sample water volume scheduling features can be implemented through the following examples.

[0069] Randomly insert interference data into the target water demand parameters of the first training instance to obtain an interference training instance, and perform feature mapping processing on the interference training instance to obtain the interference mapping feature parameters of the interference training instance;

[0070] Perform feature mapping processing on the first training instance to obtain the first mapping feature parameter instance of the first training instance;

[0071] Based on the first original abnormal event evaluation sub-model, extract the water volume scheduling vector representation of the interference mapping feature parameters to obtain the third water volume scheduling vector representation instance of the interference mapping feature parameters, and perform feature extraction operations on the third water volume scheduling vector representation instance to obtain the third sample water volume scheduling feature corresponding to the third water volume scheduling vector representation instance; the third sample water volume scheduling feature belongs to the sample water volume scheduling features;

[0072] Based on the first original abnormal event evaluation sub-model, extract the water volume scheduling vector representation of the first mapping feature parameter instance to obtain the fourth water volume scheduling vector representation instance of the first mapping feature parameter instance, and perform feature extraction operations on the fourth water volume scheduling vector representation instance to obtain the fourth sample water volume scheduling feature corresponding to the fourth water volume scheduling vector representation instance; the fourth sample water volume scheduling feature belongs to the sample water volume scheduling features;

[0073] Determine the error parameters corresponding to the interference addition training objective according to the third sample water volume scheduling feature and the fourth sample water volume scheduling feature.

[0074] In an embodiment of the present invention, by way of example, assume that the target water conservancy project is a large-scale comprehensive water conservancy hub, including multiple reservoirs, a complex irrigation canal network, numerous sluice gates, pumping stations and other facilities. The server is responsible for processing and analyzing the relevant data of the water conservancy project to ensure its normal operation and accurately evaluate abnormal events. The server selects a first training instance from the historical operation data of the water conservancy project. This first training instance includes various data related to water volume scheduling, such as reservoir water levels, river flows, sluice gate openings, and target water use demand parameters (such as specific values of agricultural irrigation water demand in a certain area, specific values of industrial water demand, etc.). Taking the agricultural irrigation water demand parameter as an example, assume that the normal irrigation water demand in a certain area within a specific time period is 5000 cubic meters per day. The server randomly inserts interference data into this target water use demand parameter according to the requirements of adding interference to the training target. For example, a random fluctuation value is added, making the irrigation water demand in this area become 6000 cubic meters per day, thus obtaining an interference training instance. Then, the server performs feature mapping processing on this interference training instance. For the reservoir water level, if its actual value is 120 meters, according to the set linear mapping rule based on the historical water level range (assuming the lowest water level is 80 meters and the highest water level is 150 meters), it is mapped to the [0, 1] interval, and the mapped value may be 0.6. Similarly, data such as river flow, sluice gate opening, and the new irrigation water demand (6000 cubic meters) are also processed according to their respective mapping rules, and finally, the interference mapping feature parameters of the interference training instance are obtained. At the same time, the server also performs feature mapping processing on the original first training instance. Still taking the above-mentioned data as an example, for the reservoir water level of 120 meters, the corresponding mapped data instance is obtained according to the same mapping rule (assuming it is also 0.6, and the actual value may vary due to different mapping rule details). Data such as river flow, sluice gate opening, and the original agricultural irrigation water demand (5000 cubic meters) are processed through their respective mapping methods respectively, so as to obtain the first mapping feature parameter instance of the first training instance. The server inputs the obtained interference mapping feature parameters into the first original abnormal event evaluation sub-model. First, the model extracts the water volume scheduling vector representation of the interference mapping feature parameters. The mapped data instances corresponding to each data are combined into a vector in a certain order to obtain the third water volume scheduling vector representation instance of the interference mapping feature parameters. For example, if after the previous processing, the mapped data instance of the reservoir water level is 0.6, the mapped data instance of the river flow is another value (assuming it is 0.4), the mapped data instance of the sluice gate opening is 0.3, and the mapped data instance of the new irrigation water demand is 0.5 (corresponding to the situation after mapping of 6000 cubic meters), then the third water volume scheduling vector representation instance can be [0.6, 0.4, 0.3, 0.5]. Then, the model performs a feature extraction operation on the third water volume scheduling vector representation instance.Extract features such as water level change trend features, flow seasonal features, sluice opening change impact features, and impact features brought about by changes in new irrigation water demand from this vector representation example, to obtain the corresponding third sample water volume scheduling features of the third water volume scheduling vector representation example. Similarly, the server inputs the first mapping feature parameter instance into the first original abnormal event evaluation sub-model. First, perform water volume scheduling vector representation extraction, and combine the mapping processing data instances corresponding to each data in a certain order to obtain the fourth water volume scheduling vector representation instance of the first mapping feature parameter instance. Assume that the mapping processing data instances of each data are similar to those before (but the mapping value corresponding to the original irrigation water demand is 5000 cubic meters), for example, [0.6, 0.4, 0.3, 0.4] (the last value here corresponds to the situation after the original irrigation water demand is mapped). Then, the model performs feature extraction operations on the fourth water volume scheduling vector representation instance, extracting features such as water level change trend features, flow seasonal features, sluice opening change impact features, and impact features related to the original irrigation water demand, to obtain the corresponding fourth sample water volume scheduling features of the fourth water volume scheduling vector representation instance. Finally, the server determines the error parameters corresponding to the interference addition training target based on the obtained third sample water volume scheduling features and fourth sample water volume scheduling features. For example, in the third sample water volume scheduling features, due to the insertion of interference data, the impact features brought about by the change in new irrigation water demand may show that it is necessary to increase the water discharge of the reservoir to meet the irrigation demand, while the fourth sample water volume scheduling features based on the original data show that the normal water discharge rhythm can meet the demand. Such differences between the two in dealing with irrigation water demand, as well as possible different manifestations in aspects such as water level, flow, and sluice opening due to changes in irrigation water demand, will be quantified into an error parameter to reflect the deviation degree of the result after the model processes the first training instance under the interference addition training target from the actual situation (i.e., the situation without adding interference data), so as to optimize and adjust the first original abnormal event evaluation sub-model subsequently.

[0075] In an embodiment of the present invention, among the multiple training targets, there is an approximate target value training target;

[0076] Processing the first training instance based on the first original abnormal event evaluation sub-model to obtain the corresponding sample water volume scheduling features of each training target, and determining the multiple error parameters corresponding to the multiple training targets according to the sample water volume scheduling features can be implemented through the following example.

[0077] Perform feature mapping processing on the first training instance to obtain the first mapping feature parameter instance of the first training instance;

[0078] Based on the first original anomaly event evaluation sub-model, extract the water volume scheduling vector representation of the first mapped feature parameter instance to obtain the fourth water volume scheduling vector representation instance of the first mapped feature parameter instance, and perform a feature extraction operation on the fourth water volume scheduling vector representation instance to obtain the corresponding fourth sample water volume scheduling feature of the fourth water volume scheduling vector representation instance; the fourth sample water volume scheduling feature belongs to the sample water volume scheduling feature;

[0079] According to the execution unit corresponding to the approximate target value training target, perform anomaly type recognition processing on the fourth sample water volume scheduling feature to obtain the training anomaly type recognition confidence corresponding to the first training instance;

[0080] Obtain the label generation model corresponding to the first original anomaly event evaluation sub-model, and perform anomaly type recognition processing on the first training instance according to the label generation model to obtain the approximate target value anomaly event prediction confidence corresponding to the first training instance;

[0081] Determine the error parameter corresponding to the approximate target value training target according to the training anomaly type recognition confidence and the approximate target value anomaly event prediction confidence.

[0082] In an embodiment of the present invention, exemplarily, the server selects a first training instance from the historical operation data of the water conservancy project, which includes a lot of water volume scheduling related information such as reservoir water level, river flow, sluice opening, pumping station operation power, and water use demands in different regions. For example, at a certain moment, the reservoir water level is 120 meters, the flow rate of a certain section of the river is 800 cubic meters per second, the sluice opening is 30%, the pumping station operation power is 250 kilowatts, and the agricultural irrigation water demand in a certain region is 5000 cubic meters per day. This instance is input into the first original abnormal event evaluation sub-model and first undergoes feature mapping processing. Taking the reservoir water level as an example, referring to the historical water level range (assuming the lowest is 80 meters and the highest is 150 meters) and the set linear mapping rule, the 120-meter water level is mapped to the interval [0,1], and the possible value is 0.6. Other data such as river flow and sluice opening are also processed according to their respective rules, and finally, the first mapped feature parameter instance of the first training instance is obtained. Then, this instance is input into the model again for extracting the water volume scheduling vector representation. The processed data instances are combined into a vector in sequence, such as [0.6, 0.4 (assuming corresponding to the mapped value of 800 cubic meters per second), 0.3, 0.25, 0.5 (assuming corresponding to the mapped value of 5000 cubic meters per day)], which is the fourth water volume scheduling vector representation instance. Subsequently, the model performs a feature extraction operation on it, and extracts features such as the water level change trend, flow seasonality, influence of sluice opening change, stability of pumping station operation power, and influence of irrigation water demand change, to obtain the fourth sample water volume scheduling features. The execution unit of the server receives these features and performs abnormal type identification processing according to specific algorithms and rules. For example, if the water level has a slow rising trend, the current dry season but the flow rate is slightly high, and the sluice adjustment frequency increases, etc., the execution unit judges whether there are abnormal types such as flood control risks and insufficient irrigation water supply based on this and the preset standards, and obtains the corresponding training abnormal type identification confidence levels. For example, the flood control risk confidence level is 0.3, and the insufficient irrigation water supply confidence level is 0.2. At the same time, the server inputs the first training instance into the label generation model trained based on a large amount of historical data and known annotation information, and it also judges based on the data features. For flood control risk, the label generation model obtains an approximate target value abnormal event prediction confidence level of 0.25; for insufficient irrigation water supply, it obtains a confidence level of 0.18. Finally, the error parameter corresponding to the approximate target value training target is determined according to the obtained training abnormal type identification confidence level and the approximate target value abnormal event prediction confidence level. Taking flood control risk as an example, the difference between the two is 0.3 - 0.25 = 0.05, which reflects the deviation degree between the model processing result and the actual situation (assuming the label generation model is closer to the actual situation). The differences of other abnormal types are also calculated in this way, and are combined to form the error parameter for subsequent optimization and adjustment of the first original abnormal event evaluation sub-model.

[0083] In the embodiment of the present invention, the target water volume scheduling feature is determined by a feature encoder in the first abnormal event evaluation sub-model performing a feature extraction operation on the water volume scheduling vector representation; the first abnormal event evaluation sub-model is a target abnormal event evaluation model for evaluating abnormal events of the target water conservancy project; the target abnormal event evaluation model further includes a second abnormal event evaluation sub-model arranged after the feature encoder;

[0084] Performing abnormal event prediction processing on the target water volume scheduling feature according to each task type to obtain the abnormal event prediction confidence corresponding to each task type can be implemented through the following examples.

[0085] Determine the second abnormal event evaluation sub-model from the target abnormal event evaluation model; the second abnormal event evaluation sub-model includes the knowledge fusion networks corresponding to each task type;

[0086] Based on the knowledge fusion networks corresponding to each task type, perform abnormal event prediction processing on the target water volume scheduling feature to obtain the abnormal event prediction confidence corresponding to each task type; one knowledge fusion network is used to determine the abnormal event prediction confidence corresponding to one task type.

[0087] In an embodiment of the present invention, exemplarily, when the server processes water conservancy project-related data, it operates according to an existing target abnormal event evaluation model. This model aims to evaluate abnormal events that may occur in water conservancy projects, and it includes a feature encoder and a second abnormal event evaluation sub-model arranged behind it. When performing abnormal event prediction processing, the server first accurately determines the second abnormal event evaluation sub-model from the target abnormal event evaluation model. Inside this sub-model, corresponding knowledge fusion networks are constructed for different task types, such as the knowledge fusion network corresponding to the flood control task type, the knowledge fusion network corresponding to the irrigation water supply task type, the knowledge fusion network corresponding to the power generation task type, etc. Taking the flood control task type as an example, the server extracts the target water volume scheduling features determined by previously performing feature extraction operations on the water volume scheduling vector representation through the feature encoder. These features may include the change trend of the reservoir water level (such as the recent water level rising rate), the seasonal fluctuation of the river flow (such as the significant increase in flow during the rainy season), the adjustment frequency of the sluice opening, etc. The server inputs these target water volume scheduling features into the knowledge fusion network corresponding to the flood control task type. This knowledge fusion network is constructed based on a large amount of historical flood control data and relevant expert knowledge. It will perform in-depth analysis and fusion processing on the input features. For example, comparing the water level rising rate with the water level rising situation during historical floods, combining the current river flow fluctuation with the flow change pattern during past flood disasters, and considering the impact of sluice opening adjustment on flood control and other factors. Through these comprehensive analyses and processes, the knowledge fusion network finally outputs the abnormal event prediction confidence corresponding to the flood control task type. Suppose after calculation and analysis, it is obtained that under the current target water volume scheduling features, the abnormal event prediction confidence for the flood control task is 0.6, that is, there is a 60% possibility of flood control-related abnormal events occurring. Similarly, for the irrigation water supply task type, the server inputs the target water volume scheduling features into its corresponding knowledge fusion network. This network will consider factors such as the matching situation between the reservoir water level and the irrigation water supply demand, the impact of sluice opening on water supply distribution, and whether the river flow can meet the irrigation demand. After processing, it outputs the abnormal event prediction confidence for the irrigation water supply task. For other task types such as the power generation task type, following a similar process, the target water volume scheduling features are input into their respective corresponding knowledge fusion networks. The knowledge fusion network analyzes and processes the target water volume scheduling features according to its own construction logic and data basis, so as to obtain the abnormal event prediction confidence corresponding to each task type, providing an important basis for the subsequent operation decision-making of the water conservancy project.

[0088] In an embodiment of the present invention, the following implementation manners are also provided.

[0089] Obtain a second sample training array; the second training instances included in the second sample training array are associated with task type target values; the task type target values include the magnitudes of the multiple task type target values corresponding to the multiple task types, one task type corresponding to one magnitude of the task type target value, and the magnitudes of the multiple task type target values being determined according to the task types corresponding to the second training instances;

[0090] Obtain the first abnormal event evaluation sub-model that has completed training. Based on the first abnormal event evaluation sub-model, perform a vector feature extraction operation on the second training instance to obtain a fifth water volume scheduling vector representation instance of the second training instance. Perform a feature extraction operation on the fifth water volume scheduling vector representation instance to obtain the fifth sample water volume scheduling feature corresponding to the fifth water volume scheduling vector representation instance;

[0091] Obtain the multiple basic knowledge fusion networks corresponding to the multiple task types. According to the multiple basic knowledge fusion networks, perform task type analysis on the fifth sample water volume scheduling feature respectively to obtain the multiple task type analysis confidence levels corresponding to the multiple basic knowledge fusion networks; one task type corresponds to one basic knowledge fusion network, and one basic knowledge fusion network is used to determine one task type analysis confidence level;

[0092] Determine a second target error parameter according to the multiple task type analysis confidence levels and the magnitudes of the multiple task type target values. According to the second target error parameter, optimize the model parameters of the multiple basic knowledge fusion networks, use the optimized multiple basic knowledge fusion networks as the multiple knowledge fusion networks, and determine the second abnormal event evaluation sub-model according to the multiple knowledge fusion networks.

[0093] In an embodiment of the present invention, exemplarily, the server carefully screens and organizes a second sample training array from the historical operation data of the water conservancy project. Each of the second training instances herein is associated with a task type target value. For example, for the flood control task type, if a certain second training instance corresponds to the water conservancy project related data (such as reservoir water level, river flow, sluice opening, etc.) before a certain flood, the magnitude of its task type target value may be set to 1, indicating that the possibility of an abnormal event (flood occurrence) in the flood control task is relatively high in this case; for the irrigation water supply task type, if another second training instance is the data during a period of tight irrigation water use, the magnitude of its task type target value may be set to 0.5, indicating that the possibility of abnormal events such as insufficient water supply in the irrigation water supply task is at a medium level. The server obtains the first abnormal event evaluation sub-model that has completed training, and sequentially inputs the second training instances in the second sample training array into this model. First, the model performs a vector feature extraction operation on the second training instance. For example, the data such as reservoir water level, river flow, sluice opening, etc. in this instance are converted into vector form according to certain rules to obtain the fifth water volume scheduling vector representation instance of the second training instance. Then, a feature extraction operation is performed on the fifth water volume scheduling vector representation instance, and the fifth sample water volume scheduling features that can reflect the characteristics of water volume scheduling, such as water level change trend, flow seasonal characteristics, sluice opening adjustment frequency, etc., are extracted from it. The server obtains multiple basic knowledge fusion networks corresponding to multiple task types (such as flood control, irrigation water supply, power generation, etc.). The extracted fifth sample water volume scheduling features are respectively input into the corresponding basic knowledge fusion networks. Taking the basic knowledge fusion network corresponding to the flood control task type as an example, it will combine historical flood control data and relevant professional knowledge to deeply analyze the input fifth sample water volume scheduling features. For example, it judges whether it conforms to the water level rising pattern before floods in history according to the water level change trend, and analyzes the possibility of flood occurrence in combination with flow characteristics, etc. After analysis, a task type analysis confidence level is output. Suppose for this flood control task type, the output task type analysis confidence level is 0.7, indicating that the model believes that the possibility of an abnormal event in the flood control task under the current input features is relatively high. Similarly, for the basic knowledge fusion networks corresponding to other task types such as irrigation water supply and power generation, they will also process the fifth sample water volume scheduling features according to their respective analysis logics and output their respective task type analysis confidence levels. The server determines the second target error parameter according to the multiple task type analysis confidence levels and the magnitudes of the corresponding multiple task type target values. For example, for the flood control task type, the task type analysis confidence level is 0.7, and the magnitude of its task type target value is 1. The difference between the two (such as by calculating the absolute value of the difference, etc.) reflects the prediction deviation degree of the current basic knowledge fusion network in this task type. By comprehensively considering this deviation situation of each task type, the second target error parameter can be determined.Then, based on the second target error parameter, the server uses a suitable optimization algorithm (such as gradient descent, etc.) to optimize the model parameters of multiple basic knowledge fusion networks. After optimization, these basic knowledge fusion networks are used as multiple knowledge fusion networks, and a second abnormal event evaluation sub-model is determined based on them, enabling it to more accurately predict abnormal events in subsequent task type analysis.

[0094] In the embodiment of the present invention, the water volume scheduling vector representation is determined by the vector representation layer in the first abnormal event evaluation sub-model extracting the water volume scheduling vector representation from the mapped feature parameters; the mapped feature parameters are determined by the data access layer in the first abnormal event evaluation sub-model performing feature mapping processing on the water volume scheduling parameters; the first abnormal event evaluation sub-model is the target abnormal event evaluation model for evaluating abnormal events of the target water conservancy project; the target abnormal event evaluation model further includes a third abnormal event evaluation sub-model arranged after the vector representation layer.

[0095] The process of extracting influence factors from the water volume scheduling vector representation to obtain the influence factor weighting coefficients corresponding to each task type can be implemented through the following examples.

[0096] Determine the third abnormal event evaluation sub-model from the target abnormal event evaluation model.

[0097] Based on the third abnormal event evaluation sub-model, perform influence factor extraction processing on the water volume scheduling vector representation to obtain the influence factor weighting coefficients corresponding to each task type.

[0098] In an embodiment of the present invention, exemplarily, when processing water conservancy project data, the server works based on an existing target abnormal event evaluation model. This model aims to comprehensively evaluate various abnormal events that may occur in water conservancy projects, including multiple components such as a data access layer, a vector representation layer, and a third abnormal event evaluation sub-model arranged after the vector representation layer. When it is necessary to perform influence factor extraction processing on the water volume scheduling vector representation, the server first accurately determines the third abnormal event evaluation sub-model from the target abnormal event evaluation model. This sub-model is a key part specifically used to analyze the influence degree of various factors on water volume scheduling under different task types, and then determine the influence factor weighting coefficient. Taking the flood control task type as an example, the server obtains the water volume scheduling vector representation obtained through a series of previous processes. This vector representation contains a lot of information related to water volume scheduling, such as the vector value corresponding to the reservoir water level, the vector value corresponding to the river flow rate, the vector value corresponding to the sluice opening degree, and the vector value corresponding to the pumping station operating power, etc. The server inputs this water volume scheduling vector representation into the third abnormal event evaluation sub-model. Based on a large amount of historical flood control data and relevant professional knowledge, this sub-model will conduct in-depth analysis on the input vector representation to extract the influence factor weighting coefficient. For the factor of reservoir water level, the model will analyze its importance in flood control tasks. For example, if historical data shows that whenever the reservoir water level rapidly rises above a certain critical value, the risk of flood occurrence will increase significantly. Then, in the current water volume scheduling vector representation, the water level situation reflected by the vector value corresponding to the reservoir water level will be given a relatively high influence factor weighting coefficient. Suppose that according to the analysis, the influence factor weighting coefficient of the reservoir water level under the flood control task type is determined to be 0.5. For the river flow rate, in the flood control scenario, the magnitude and change trend of the flow rate have an important impact on the formation and development of floods. When the river flow rate continues to increase and exceeds the normal flood discharge capacity range, it is easy to trigger floods. Therefore, according to the vector value corresponding to the river flow rate in the current water volume scheduling vector representation and the comparison and analysis of historical data, the influence factor weighting coefficient of the river flow rate under the flood control task type may be determined to be 0.3. Similarly, for factors such as the sluice opening degree and the pumping station operating power, the third abnormal event evaluation sub-model will also analyze them according to their actual influence in flood control tasks. For example, the reasonable adjustment of the sluice opening degree is crucial for controlling the reservoir flood discharge volume and regulating the river flow rate to cope with floods. If the current opening degree situation shows that it has a certain impact on flood control, after analysis, it may be given an influence factor weighting coefficient of 0.1; the influence of the pumping station operating power in flood control tasks is relatively small, and its influence factor weighting coefficient may be determined to be 0.1.Similarly, for other task types such as irrigation water supply task type and power generation task type, the server will also input the corresponding water volume scheduling vector representation into the third abnormal event evaluation sub-model. According to the characteristics and requirements of each task, by analyzing the importance of each factor in the task, the influence factor weighting coefficients corresponding to each factor under each task type are determined, so as to comprehensively and accurately evaluate the influence degree of each factor on water volume scheduling under different task types.

[0099] In the embodiment of the present invention, the target abnormal event evaluation model further includes a second abnormal event evaluation sub-model for determining the prediction confidence of abnormal events corresponding to each task type; the embodiment of the present invention also provides the following implementation manners.

[0100] Obtain a third sample training array; the third training instances included in the third sample training array are associated with abnormal event target values;

[0101] Obtain the first abnormal event evaluation sub-model that has completed training. Based on the first abnormal event evaluation sub-model, perform a vector feature extraction operation on the third training instance to obtain the sixth water volume scheduling vector representation instance of the third training instance. Perform a feature extraction operation on the sixth water volume scheduling vector representation instance to obtain the corresponding sixth sample water volume scheduling feature of the sixth water volume scheduling vector representation instance;

[0102] Obtain the second abnormal event evaluation sub-model that has completed training. Based on multiple knowledge fusion networks in the second abnormal event evaluation sub-model, perform abnormal event prediction processing on the sixth sample water volume scheduling feature to obtain the training abnormal event prediction confidence corresponding to each task type; one knowledge fusion network is used to determine the training abnormal event prediction confidence corresponding to one task type;

[0103] Based on the third original abnormal event evaluation sub-model, perform influence factor extraction processing on the sixth water volume scheduling vector representation to obtain the training influence factor weighting coefficients corresponding to each task type;

[0104] According to the training abnormal event prediction confidence corresponding to each task type and the training influence factor weighting coefficients corresponding to each task type, determine the training abnormal event evaluation confidence of the virtual water conservancy project corresponding to the third training instance;

[0105] Determine a third target error parameter according to the abnormal event target value and the training abnormal event evaluation confidence level. Optimize the model parameters of the third original abnormal event evaluation sub-model according to the third target error parameter, and use the optimized third original abnormal event evaluation sub-model as the third abnormal event evaluation sub-model. Determine the target abnormal event evaluation model according to the first abnormal event evaluation sub-model, the second abnormal event evaluation sub-model and the third abnormal event evaluation sub-model.

[0106] In an embodiment of the present invention, exemplarily, the server carefully selects and sorts out a third sample training array from the massive historical operation data of the water conservancy project, where each third training instance is associated with an abnormal event target value. For the flood control task, if a certain third training instance corresponds to the relevant data during a certain flood occurrence, such as the reservoir water level, river flow rate, sluice opening degree, etc., its abnormal event target value is set to 1, indicating that a serious abnormality (flood occurrence) has occurred in the flood control task; for the irrigation water supply task, if the instance is the data during the period of insufficient irrigation water supply, the abnormal event target value may be set to 0.5, indicating that a relatively obvious abnormality has occurred. The server first obtains the first abnormal event evaluation sub-model that has completed training and sequentially inputs the third training instances in the third sample training array. The model first performs a vector feature extraction operation on the instance, converts data such as the reservoir water level, river flow rate, and sluice opening degree into vector form according to specific rules, and obtains the sixth water volume scheduling vector representation instance. Then, it extracts the sixth sample water volume scheduling features that can reflect the characteristics of water volume scheduling, such as the water level change trend, flow seasonal characteristics, sluice opening degree adjustment frequency, etc. Subsequently, the server obtains the second abnormal event evaluation sub-model that has completed training, and this model contains multiple knowledge fusion networks corresponding to different task types. The extracted sixth sample water volume scheduling features are respectively input into the corresponding knowledge fusion networks. Taking the knowledge fusion network corresponding to the flood control task as an example, it will combine historical flood control data and professional knowledge to deeply analyze the input features. For example, it judges whether it conforms to the water level rising pattern before the historical flood comes based on the water level change trend, and analyzes the possibility of flood occurrence by combining the flow characteristics, etc., and then outputs the training abnormal event prediction confidence of the flood control task type. Suppose it is 0.8, that is, the model believes that the possibility of an abnormality occurring in the flood control task under the current input features is relatively high. The knowledge fusion networks of other task types also process the features according to their respective logics and output the corresponding training abnormal event prediction confidences. The server performs influence factor extraction processing on the sixth water volume scheduling vector representation based on the third original abnormal event evaluation sub-model. Taking the irrigation water supply task as an example, it analyzes the importance of each factor in the task. The reservoir water level is crucial. Too low water level may lead to insufficient water supply, and its training influence factor weighting coefficient may be 0.4; reasonable adjustment of the sluice opening degree affects water supply distribution, and the coefficient may be 0.3; the river flow rate is related to the water supply situation, and the coefficient may be 0.2; the operation power of the pumping station has a relatively small impact, and the coefficient may be 0.1. The same principle is used to determine the respective coefficients for other task types. Then, according to the training abnormal event prediction confidences and training influence factor weighting coefficients of each task type, the training abnormal event evaluation confidence of the virtual water conservancy project corresponding to the third training instance is determined. For the flood control task type, given the relevant data, the contribution value to the overall evaluation confidence is obtained through weighted calculation, and then combined with the contribution values of other task types to obtain the training abnormal event evaluation confidence of this instance.Then, determine the third target error parameter based on the abnormal event target value of the third training instance and the confidence level of the training abnormal event evaluation. For example, the difference between the two can be part of it, and determine the final parameter by comprehensively considering the situations of each task type. Finally, according to the third target error parameter, the server uses a suitable optimization algorithm (such as gradient descent, etc.) to optimize the parameters of the third original abnormal event evaluation sub-model, and uses the optimized model as the third abnormal event evaluation sub-model. Then, combined with the already trained first and second abnormal event evaluation sub-models, determine the complete target abnormal event evaluation model to make it play a more accurate role in the subsequent evaluation of abnormal events in water conservancy projects.

[0107] In an embodiment of the present invention, the third abnormal event evaluation sub-model includes a decision-making component and a weight allocation component;

[0108] Perform impact factor extraction processing on the water volume scheduling vector representation based on the third abnormal event evaluation sub-model to obtain the impact factor weighting coefficients corresponding to each task type, which can be implemented through the following examples.

[0109] Perform decision feature extraction operations on the water volume scheduling vector representation based on the decision-making component to obtain the decision extraction features corresponding to the water volume scheduling vector representation;

[0110] Perform impact factor identification on the decision extraction features based on the weight allocation component to obtain the impact factor weighting coefficients corresponding to each task type.

[0111] In an embodiment of the present invention, by way of example, the server obtains the pre-processed water volume scheduling vector representation, including relevant information such as reservoir water level, river flow, sluice opening, and pump station operating power, and inputs it into the decision-making component of the third abnormal event evaluation sub-model. The decision-making component deeply analyzes according to the built-in rule algorithm and extracts decision-making significant features. Taking the flood control task as an example, it focuses on features such as the rising trend of the reservoir water level, abnormal changes in river flow, and abnormal adjustments in sluice opening. The server then inputs the decision extraction features into the weight allocation component, which performs impact factor identification according to the preset rules, historical data, and professional knowledge to determine the impact factor weighting coefficients for each task type. For example, in the flood control task, the weighting coefficient of the rising trend feature of the water level may be 0.5, the abnormal change feature of the flow may be 0.3, and the abnormal adjustment feature of the sluice opening may be 0.1; in the irrigation water supply task, the reservoir water level, sluice opening, river flow, pump station operating power, etc. correspond to different weighting coefficients. In this way, the weight allocation component can provide a basis for evaluating the status of water conservancy projects.

[0112] In an embodiment of the present invention, the multiple task types include the target task type; the embodiment of the present invention also provides the following implementation methods.

[0113] Obtain a target abnormal event evaluation model for evaluating abnormal events of a target water conservancy project, and obtain a basic integrated model for transfer learning from the target abnormal event evaluation model;

[0114] Obtain a fourth sample training array corresponding to the target task type; the fourth sample training array includes fourth training instances;

[0115] Process the fourth training instance based on the target abnormal event evaluation model to obtain the abnormal event evaluation confidence corresponding to the fourth training instance, and use the abnormal event evaluation confidence corresponding to the fourth training instance as an approximate target value instance for training the basic integrated model;

[0116] Process the fourth training instance based on the basic integrated model to obtain the integrated abnormal event evaluation confidence corresponding to the fourth training instance;

[0117] Optimize the model parameters of the basic integrated model according to the integrated abnormal event evaluation confidence and the approximate target value instance, and use the optimized basic integrated model as the target integrated model of the target abnormal event evaluation model; the target integrated model is used to evaluate abnormal events of water volume scheduling parameters under the target task type.

[0118] In an embodiment of the present invention, exemplarily, the server first obtains a target abnormal event evaluation model for evaluating abnormal events of a target water conservancy project. This model has been constructed and trained in the early stage and already has a certain evaluation ability. At the same time, the server obtains a basic integrated model for transfer learning from the target abnormal event evaluation model. This basic integrated model can be trained based on data from other similar water conservancy projects or different stages of the same water conservancy project, and has a certain generality but still needs to be further optimized for the current target task type. Then, the server obtains a fourth sample training array corresponding to the target task type, which contains multiple fourth training instances. Taking the irrigation water supply task type as an example of the target task type, the fourth training instances can be water conservancy project data records related to irrigation water supply at different time periods and seasons, such as specific data situations of reservoir water levels, river flows, sluice openings, and water use demands in the irrigation area during a certain period. The server inputs the fourth training instances in the fourth sample training array into the target abnormal event evaluation model one by one for processing. Still taking the irrigation water supply task type as an example, for a specific fourth training instance, assuming it records data for a certain period in summer, the reservoir water level is 120 meters, the river flow is 500 cubic meters per second, the sluice opening is 30%, and the water use demand in the irrigation area is 3000 cubic meters per day. The target abnormal event evaluation model will analyze and evaluate these input data according to its internal algorithms and existing data patterns, and obtain the abnormal event evaluation confidence corresponding to this fourth training instance. For example, after evaluation, it is considered that the abnormal event evaluation confidence of insufficient irrigation water supply in the current data situation is 0.3, that is, there is a 30% possibility of insufficient irrigation water supply. Then, the server uses this abnormal event evaluation confidence as an approximate target value instance for training the basic integrated model. The server then inputs the same fourth training instance into the basic integrated model for processing. The basic integrated model analyzes the input fourth training instance data based on its own structure and algorithms. For the fourth training instance of the above irrigation water supply task type, the basic integrated model may comprehensively consider factors such as reservoir water level, river flow, sluice opening, and water use demand, and through its internal integration mechanism (such as the comprehensive judgment of multiple sub-models, etc.), obtain the integrated abnormal event evaluation confidence corresponding to this fourth training instance. Assuming that after processing, the obtained integrated abnormal event evaluation confidence is 0.25, that is, there is a 25% possibility of abnormal events related to irrigation water supply. The server optimizes the model parameters of the basic integrated model according to the obtained integrated abnormal event evaluation confidence and the approximate target value instance. Continuing to take the irrigation water supply task type as an example, the approximate target value instance is 0.3 (from the evaluation result of the target abnormal event evaluation model), and the integrated abnormal event evaluation confidence is 0.25 (from the processing result of the basic integrated model).There are certain differences between the two, indicating that there is a deviation between the evaluation results of the basic integration model and the target abnormal event evaluation model. The server adjusts the model parameters of the basic integration model based on this deviation through specific optimization algorithms (such as gradient descent, etc.) to make its evaluation results closer to those of the target abnormal event evaluation model. After multiple rounds of such optimization processes, the optimized basic integration model is used as the target integration model of the target abnormal event evaluation model. This target integration model is specifically used to evaluate abnormal events for water volume scheduling parameters under the target task type (such as the irrigation water supply task type), and can more accurately judge the possible abnormal event situations under this task type, providing a more reliable basis for the reasonable scheduling and operation of water conservancy projects.

[0119] In the embodiment of the present invention, the fourth training instance includes task training instances generated under the target task type;

[0120] The obtaining of the fourth sample training array corresponding to the target task type can be implemented through the following examples.

[0121] Obtain task training instances under the target task type;

[0122] Determine pending training instances from the training instance set for instance selection;

[0123] Based on the training instance expansion model, match and associate the pending training instances with the task training instances to obtain the matching coefficients between the task training instances and the pending training instances;

[0124] If the matching coefficient meets the instance expansion condition, then use the pending training instance as the expanded training instance corresponding to the target task type;

[0125] Use the expanded training instance and the task training instance as the fourth training instance corresponding to the target task type.

[0126] In an embodiment of the present invention, exemplarily, the server first specifically selects task training instances related to the irrigation water supply task type from the historical operation data of the water conservancy project. These task training instances record the data conditions of each key link in the actual irrigation water supply process. For example, during the peak spring irrigation period, there is a task training instance that records the water level of a certain reservoir as 110 meters. At this time, the river flow rate at the entrance of the irrigation canal is 400 cubic meters per second, the corresponding sluice opening is set at 40%, and the water demand in the irrigation area is 2500 cubic meters per day. Such detailed data records can reflect the actual situation of irrigation water supply at that time, providing a real and targeted data basis for subsequent model training. The server then determines the pending training instances from the training instance set used for instance selection. This training instance set contains various data records of the water conservancy project under different operation scenarios and different time periods. For example, a data record that selects a certain period with a reservoir water level of 105 meters, a river flow rate of 350 cubic meters per second at a certain section, a sluice opening of 35%, and some other water demand situations (such as industrial water use, etc.) in the surrounding area during that period is selected as a pending training instance. Although this instance is not entirely for the irrigation water supply task type, it contains some factors that may be related to irrigation water supply, so it is selected as an object to be further analyzed. The server inputs the obtained pending training instances and the task training instances under the irrigation water supply task type previously selected into the training instance expansion model. This training instance expansion model will comprehensively analyze these two types of instances according to a set of rules and algorithms to determine the matching coefficient between them. For example, for the above-mentioned task training instance (the situation during the peak spring irrigation period) and the pending training instance (the situation with various water demand situations during a certain period), the model will compare their reservoir water level data, analyze the potential impact of the water level difference and the water level change trend on irrigation water supply; at the same time, compare the river flow rate data, consider the role of the flow rate magnitude and fluctuation on irrigation water supply; and also pay attention to the similarities and differences in the sluice opening and the impact of other water demand situations on the allocation of irrigation water supply resources and other aspects of factors. After complex calculations and analyses, assume that the matching coefficient between these two instances is 0.6. This matching coefficient reflects the similarity degree of the pending training instance and the task training instance in terms of factors related to the irrigation water supply task. The server then determines whether this matching coefficient meets the instance expansion condition. Assume that the set instance expansion condition is that the matching coefficient is greater than or equal to 0.5. Since the previously obtained matching coefficient is 0.6, which meets the instance expansion condition, the server uses the pending training instance as the expanded training instance corresponding to the target task type (irrigation water supply task type). Finally, the server uses the expanded training instance and the original task training instances together as the fourth training instance corresponding to the target task type.The fourth set of training instances thus formed contains both actual data records (task training instances) under the pure irrigation water supply task type and other data records (augmented training instances) with a certain degree of relevance after screening and matching. It can provide a richer and more comprehensive data basis for subsequent model training based on the target task type, and helps to improve the accuracy of the model's assessment of abnormal events under the irrigation water supply task type.

[0127] In the embodiment of the present invention, to determine the abnormal event assessment confidence of the target water conservancy project according to the abnormal event prediction confidence corresponding to each task type and the influence factor weighting coefficient corresponding to each task type, the following example can be executed.

[0128] Weight the abnormal event prediction confidence corresponding to each task type based on the influence factor weighting coefficient corresponding to each task type to obtain the weighted abnormal event prediction coefficient corresponding to each task type; the influence factor conversion value corresponding to a task type is used to weight the abnormal event prediction confidence corresponding to the corresponding task type.

[0129] Determine the abnormal event assessment confidence of the target water conservancy project according to the weighted abnormal event prediction coefficient corresponding to each task type.

[0130] In the embodiment of the present invention, by way of example, in this water conservancy project, multiple task types such as flood control, irrigation water supply, and power generation are set. The server has obtained the abnormal event prediction confidence and influence factor weighting coefficients of each task type through models and algorithms. Taking the flood control task as an example, its abnormal event prediction confidence is 0.6, and the influence factor weighting coefficients of factors such as reservoir water level, river flow, sluice opening, and pump station operating power are 0.4, 0.3, 0.2, and 0.1 respectively. After weighted calculation, the weighted abnormal event prediction coefficient is 0.6. Similarly, for the irrigation water supply task type, assuming the abnormal event prediction confidence is 0.4 and the weighting coefficients of each factor are different, the weighted abnormal event prediction coefficient is 0.4; the power generation task type and others are also calculated to obtain the corresponding coefficients in this way. After that, when the server needs to determine the abnormal event assessment confidence of the target water conservancy project, methods such as weighted average or simple average can be used. If only considering the above three task types and knowing their weighted abnormal event prediction coefficients, the abnormal event assessment confidence obtained by simple average is approximately 0.43, and the weighted average (assuming the weights of each task type) is 0.45. Through these calculations, comprehensively considering the abnormal predictions of each task type and the importance of factors, the overall abnormal event assessment confidence of the target water conservancy project is obtained, which is used to evaluate and make decisions on the overall operation status of the project.

[0131] An embodiment of the present invention provides a computer device 100. The computer device 100 includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the foregoing method for optimizing operation parameters and data governance in water conservancy project construction. As Figure 2 shown, Figure 2 is a structural block diagram of the computer device 100 provided by an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To achieve data transmission or interaction, the elements of the memory 111, the processor 112, and the communication unit 113 are electrically connected to each other directly or indirectly. For example, these elements can be electrically connected to each other through one or more communication buses or signal lines.

[0132] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the disclosure to the precise forms disclosed. According to the above teachings, numerous modifications and variations are possible. These embodiments are selected and described to best illustrate the principles of the disclosure and its practical applications, so that those skilled in the art can best utilize the disclosure and use various embodiments with different modifications to suit the particular applications contemplated.

Claims

1. A method for optimizing operation parameters and data governance in water conservancy project construction, characterized in that Including: Performing a vector feature extraction operation on the water volume scheduling parameters of the target water conservancy project to obtain a water volume scheduling vector representation corresponding to the water volume scheduling parameters; Performing a feature extraction operation on the water volume scheduling vector representation to obtain target water volume scheduling features corresponding to the water volume scheduling vector representation; Determining multiple task types corresponding to the water volume scheduling parameters, and respectively performing abnormal event prediction processing on the target water volume scheduling features according to each task type to obtain abnormal event prediction confidence levels corresponding to each task type; Performing influence factor extraction processing on the water volume scheduling vector representation to obtain influence factor weighting coefficients corresponding to each task type; Weighting the abnormal event prediction confidence levels corresponding to each task type based on the influence factor weighting coefficients corresponding to each task type to obtain weighted abnormal event prediction coefficients corresponding to each task type; An influence factor conversion value corresponding to a task type is used to weight the abnormal event prediction confidence level corresponding to the corresponding task type; Determining the abnormal event evaluation confidence level of the target water conservancy project according to the weighted abnormal event prediction coefficients corresponding to each task type; When the abnormal event evaluation confidence level indicates that there is an abnormal event in the target water conservancy project, collecting real-time monitoring data of each subsystem and key equipment of the target water conservancy project, and combining the abnormal event evaluation confidence level to push a preset governance strategy to each subsystem for optimization and adjustment; The performing a vector feature extraction operation on the water volume scheduling parameters of the target water conservancy project to obtain a water volume scheduling vector representation corresponding to the water volume scheduling parameters includes: When obtaining the water volume scheduling parameters of the target water conservancy project, obtaining a target abnormal event evaluation model for evaluating abnormal events of the target water conservancy project; The target abnormal event evaluation model includes a first abnormal event evaluation submodel; Determining demand information corresponding to the water volume scheduling parameters based on the first abnormal event evaluation submodel, and extracting target water use demand parameters of the water volume scheduling parameters according to the demand information; The first abnormal event evaluation submodel includes a data access layer; The target water use demand parameters include multiple target water use demand parameters, and the multiple target water use demand parameters include target key water use demand parameters; Determining a transform domain corresponding to the target key water use demand parameter according to the demand information corresponding to the target key water use demand parameter; The transform domain includes multiple conversion values, and each conversion value has a corresponding target water use demand numerical range; Determining the target water use demand numerical range corresponding to the target key water use demand parameter based on the data access layer, and using the conversion value corresponding to the target water use demand numerical range corresponding to the target key water use demand parameter as the mapping processing data corresponding to the target key water use demand parameter; Performing a normalization operation on the mapping processing data corresponding to the target water use demand parameters based on the first abnormal event evaluation submodel to obtain the mapping feature parameters of the water volume scheduling parameters; Extract the water volume scheduling vector representation of the mapping feature parameters to obtain the water volume scheduling vector representation of the mapping feature parameters.

2. The method according to claim 1, wherein The target water volume scheduling feature is determined by the feature encoder in the first abnormal event evaluation sub-model performing a feature extraction operation on the water volume scheduling vector representation; the first abnormal event evaluation sub-model is the target abnormal event evaluation model for evaluating abnormal events of the target water conservancy project; The method further includes: Obtain a first sample training array; the first training instances included in the first sample training array are training instances without configured abnormal event target values; Obtain multiple training targets for the first original abnormal event evaluation sub-model; Based on the first original abnormal event evaluation sub-model, process the first training instance to obtain sample water volume scheduling features corresponding to each training target, and determine multiple error parameters corresponding to the multiple training targets according to the sample water volume scheduling features; one training target corresponds to one error parameter; Determine a first target error parameter according to the multiple error parameters, optimize the model parameters of the first original abnormal event evaluation sub-model according to the first target error parameter, and use the optimized first original abnormal event evaluation sub-model as the first abnormal event evaluation sub-model.

3. The method according to claim 2, characterized in that, The multiple training targets include a random deletion training target; The processing the first training instance based on the first original abnormal event evaluation sub-model to obtain sample water volume scheduling features corresponding to each training target, and determining multiple error parameters corresponding to the multiple training targets according to the sample water volume scheduling features includes: Perform feature mapping processing on the first training instance based on the first original abnormal event evaluation sub-model to obtain a first mapping feature parameter instance of the first training instance; the first mapping feature parameter instance includes mapping processing data instances on multiple index tags; Randomly select a first index tag to be randomly deleted from the multiple index tags corresponding to the first mapping feature parameter instance; Perform random deletion processing on the mapping processing data instance on the first index tag in the first mapping feature parameter instance according to the random deletion information to obtain a first feature parameter instance; Extract the water volume scheduling vector representation of the first feature parameter instance based on the first original abnormal event evaluation sub-model to obtain a first water volume scheduling vector representation instance of the first feature parameter instance, and perform a feature extraction operation on the first water volume scheduling vector representation instance to obtain a first sample water volume scheduling feature corresponding to the first water volume scheduling vector representation instance; the first sample water volume scheduling feature belongs to the sample water volume scheduling features; Based on the execution unit corresponding to the random deletion training target, perform random deletion inference processing on the first sample water volume scheduling feature to obtain the inference mapping processing data corresponding to the random deletion information in the first feature parameter instance; Determine the error parameter corresponding to the random deletion training target according to the mapping processing data instance on the first index tag in the first mapping feature parameter instance and the inferred mapping processing data; The multiple training targets further include a local permutation training target; Processing the first training instance based on the first original abnormal event evaluation sub-model to obtain the sample water volume scheduling features corresponding to each training target, and determining the multiple error parameters corresponding to the multiple training targets according to the sample water volume scheduling features, including: Performing feature mapping processing on the first training instance based on the first original abnormal event evaluation sub-model to obtain the first mapping feature parameter instance of the first training instance; the first mapping feature parameter instance includes mapping processing data instances on multiple index tags; Randomly select a second index tag to be locally permuted from the multiple index tags corresponding to the first mapping feature parameter instance; Perform local permutation processing on the mapping processing data instance on the second index tag in the first mapping feature parameter instance according to the preset element information to obtain a second feature parameter instance; the preset element information is different from the mapping processing data instance on the second index tag; Extract the water volume scheduling vector representation of the second feature parameter instance based on the first original abnormal event evaluation sub-model, obtain the second water volume scheduling vector representation instance of the second feature parameter instance, and perform a feature extraction operation on the second water volume scheduling vector representation instance to obtain the second sample water volume scheduling feature corresponding to the second water volume scheduling vector representation instance; the second sample water volume scheduling feature belongs to the sample water volume scheduling features; Perform local permutation inference processing on the second sample water volume scheduling feature based on the execution unit corresponding to the local permutation training target to obtain the local permutation inference results corresponding to the multiple index tags; Determine the error parameter corresponding to the local permutation training target according to the second index tag and the local permutation inference results corresponding to the multiple index tags; The multiple training targets further include an interference addition training target; Processing the first training instance based on the first original abnormal event evaluation sub-model to obtain the sample water volume scheduling features corresponding to each training target, and determining the multiple error parameters corresponding to the multiple training targets according to the sample water volume scheduling features, including: Randomly insert interference data into the target water demand parameter of the first training instance to obtain an interference training instance, and perform feature mapping processing on the interference training instance to obtain the interference mapping feature parameter of the interference training instance; Perform feature mapping processing on the first training instance to obtain the first mapping feature parameter instance of the first training instance; Based on the first original abnormal event evaluation sub-model, extract the water volume scheduling vector representation of the interference mapping feature parameters to obtain the third water volume scheduling vector representation instance of the interference mapping feature parameters, and perform a feature extraction operation on the third water volume scheduling vector representation instance to obtain the corresponding third sample water volume scheduling feature of the third water volume scheduling vector representation instance; the third sample water volume scheduling feature belongs to the sample water volume scheduling features; Based on the first original abnormal event evaluation sub-model, extract the water volume scheduling vector representation of the first mapping feature parameter instance to obtain the fourth water volume scheduling vector representation instance of the first mapping feature parameter instance, and perform a feature extraction operation on the fourth water volume scheduling vector representation instance to obtain the corresponding fourth sample water volume scheduling feature of the fourth water volume scheduling vector representation instance; the fourth sample water volume scheduling feature belongs to the sample water volume scheduling features; According to the third sample water volume scheduling feature and the fourth sample water volume scheduling feature, determine the error parameter corresponding to the interference addition training target; Among the multiple training targets, there is also an approximate target value training target; Processing the first training instance based on the first original abnormal event evaluation sub-model to obtain the sample water volume scheduling features corresponding to each training target, and determining the multiple error parameters corresponding to the multiple training targets according to the sample water volume scheduling features, including: Perform a feature mapping process on the first training instance to obtain the first mapping feature parameter instance of the first training instance; Based on the first original abnormal event evaluation sub-model, extract the water volume scheduling vector representation of the first mapping feature parameter instance to obtain the fourth water volume scheduling vector representation instance of the first mapping feature parameter instance, and perform a feature extraction operation on the fourth water volume scheduling vector representation instance to obtain the corresponding fourth sample water volume scheduling feature of the fourth water volume scheduling vector representation instance; the fourth sample water volume scheduling feature belongs to the sample water volume scheduling features; According to the execution unit corresponding to the approximate target value training target, perform an abnormal type recognition process on the fourth sample water volume scheduling feature to obtain the training abnormal type recognition confidence corresponding to the first training instance; Obtain a label generation model corresponding to the first original abnormal event evaluation sub-model, and perform an abnormal type recognition process on the first training instance according to the label generation model to obtain the approximate target value abnormal event prediction confidence corresponding to the first training instance; According to the training abnormal type recognition confidence and the approximate target value abnormal event prediction confidence, determine the error parameter corresponding to the approximate target value training target.

4. The method according to claim 1, wherein The target water volume scheduling feature is determined by the feature encoder in the first abnormal event evaluation sub-model performing a feature extraction operation on the water volume scheduling vector representation; the first abnormal event evaluation sub-model is the target abnormal event evaluation model for evaluating abnormal events of the target water conservancy project; the target abnormal event evaluation model also includes a second abnormal event evaluation sub-model arranged after the feature encoder; Performing anomaly event prediction processing on the target water volume scheduling features according to each task type respectively to obtain the anomaly event prediction confidence levels corresponding to each task type, including: Determining a second anomaly event evaluation sub-model from the target anomaly event evaluation model; the second anomaly event evaluation sub-model includes the knowledge fusion networks corresponding to each task type; Based on the knowledge fusion networks corresponding to each task type, performing anomaly event prediction processing on the target water volume scheduling features to obtain the anomaly event prediction confidence levels corresponding to each task type; one knowledge fusion network is used to determine the anomaly event prediction confidence level corresponding to one task type.

5. The method according to claim 4, wherein The method further includes: Obtaining a second sample training array; the second training instances included in the second sample training array are associated with task type target values; the task type target values include the magnitudes of the multiple task type target values corresponding to the multiple task types, one task type corresponds to the magnitude of one task type target value, and the magnitudes of the multiple task type target values are determined according to the task type corresponding to the second training instance; Obtaining the first anomaly event evaluation sub-model that has completed training, performing a vector feature extraction operation on the second training instance based on the first anomaly event evaluation sub-model to obtain the fifth water volume scheduling vector representation instance of the second training instance, and performing a feature extraction operation on the fifth water volume scheduling vector representation instance to obtain the fifth sample water volume scheduling features corresponding to the fifth water volume scheduling vector representation instance; Obtaining the multiple basic knowledge fusion networks corresponding to the multiple task types, performing task type analysis on the fifth sample water volume scheduling features according to the multiple basic knowledge fusion networks respectively to obtain the multiple task type analysis confidence levels corresponding to the multiple basic knowledge fusion networks; one task type corresponds to one basic knowledge fusion network, and one basic knowledge fusion network is used to determine one task type analysis confidence level; Determining a second target error parameter according to the multiple task type analysis confidence levels and the magnitudes of the multiple task type target values, optimizing the model parameters of the multiple basic knowledge fusion networks according to the second target error parameter, using the optimized multiple basic knowledge fusion networks as the multiple knowledge fusion networks, and determining the second anomaly event evaluation sub-model according to the multiple knowledge fusion networks.

6. The method according to claim 1, wherein The water volume scheduling vector representation is determined by the vector representation layer in the first anomaly event evaluation sub-model extracting the water volume scheduling vector representation from the mapped feature parameters; the mapped feature parameters are determined by the data access layer in the first anomaly event evaluation sub-model performing feature mapping processing on the water volume scheduling parameters; the first anomaly event evaluation sub-model is the target anomaly event evaluation model for evaluating anomaly events of the target water conservancy project; the target anomaly event evaluation model further includes a third anomaly event evaluation sub-model arranged after the vector representation layer; Performing influence factor extraction processing on the water volume scheduling vector representation to obtain the influence factor weighting coefficients corresponding to each task type, including: Determine a third abnormal event evaluation sub-model from the target abnormal event evaluation model; the third abnormal event evaluation sub-model includes a decision-making component and a weight allocation component; Perform a decision feature extraction operation on the water volume scheduling vector representation based on the decision-making component to obtain the corresponding decision extraction features of the water volume scheduling vector representation; Identify the influence factor weighting coefficients corresponding to each task type based on the weight allocation component for the decision extraction features; 7. The method according to claim 6, characterized in that, The target abnormal event evaluation model further includes a second abnormal event evaluation sub-model for determining the abnormal event prediction confidence corresponding to each task type; The method further includes: Obtain a third sample training array; the third training instances included in the third sample training array are associated with abnormal event target values; Obtain the first abnormal event evaluation sub-model that has completed training, perform a vector feature extraction operation on the third training instance based on the first abnormal event evaluation sub-model to obtain the sixth water volume scheduling vector representation instance of the third training instance, and perform a feature extraction operation on the sixth water volume scheduling vector representation instance to obtain the corresponding sixth sample water volume scheduling features of the sixth water volume scheduling vector representation instance; Obtain the second abnormal event evaluation sub-model that has completed training, perform abnormal event prediction processing on the sixth sample water volume scheduling features based on multiple knowledge fusion networks in the second abnormal event evaluation sub-model to obtain the corresponding training abnormal event prediction confidences for each task type; one knowledge fusion network is used to determine the training abnormal event prediction confidence corresponding to one task type; Perform influence factor extraction processing on the sixth water volume scheduling vector representation based on the third original abnormal event evaluation sub-model to obtain the corresponding training influence factor weighting coefficients for each task type; Determine the training abnormal event evaluation confidence of the virtual water conservancy project corresponding to the third training instance according to the training abnormal event prediction confidence corresponding to each task type and the training influence factor weighting coefficients corresponding to each task type; Determine a third target error parameter according to the abnormal event target value and the training abnormal event evaluation confidence, optimize the model parameters of the third original abnormal event evaluation sub-model according to the third target error parameter, use the optimized third original abnormal event evaluation sub-model as the third abnormal event evaluation sub-model, and determine the target abnormal event evaluation model according to the first abnormal event evaluation sub-model, the second abnormal event evaluation sub-model, and the third abnormal event evaluation sub-model.

8. The method according to claim 1, wherein The multiple task types include a target task type; The method further includes: Obtain a target abnormal event evaluation model for evaluating abnormal events of a target water conservancy project, and obtain a basic integration model for transfer learning from the target abnormal event evaluation model; Obtain the task training instances under the target task type; Determine the pending training instances from the training instance set for instance selection; Based on the training instance expansion model, match and associate the to-be-determined training instance with the task training instance to obtain the matching coefficient between the task training instance and the to-be-determined training instance; If the matching coefficient meets the instance expansion condition, use the to-be-determined training instance as the expanded training instance corresponding to the target task type; Use the expanded training instance and the task training instance as the fourth training instance corresponding to the target task type; the fourth training instance includes the task training instances generated under the target task type; Based on the target abnormal event evaluation model, process the fourth training instance to obtain the abnormal event evaluation confidence corresponding to the fourth training instance, and use the abnormal event evaluation confidence corresponding to the fourth training instance as the approximate target value instance for training the basic integration model; Based on the basic integration model, process the fourth training instance to obtain the integrated abnormal event evaluation confidence corresponding to the fourth training instance; Optimize the model parameters of the basic integration model according to the integrated abnormal event evaluation confidence and the approximate target value instance, and use the optimized basic integration model as the target integration model of the target abnormal event evaluation model; the target integration model is used to evaluate abnormal events for the water volume scheduling parameters under the target task type.

9. A server system, characterized in that, It includes a server, and the server is used to execute the method described in any one of claims 1-8.

Images