A multi-model fusion multi-reservoir area general inflow forecasting system

Abstract

The application relates to the technical field of hydrological prediction, and particularly discloses a multi-model fusion multi-reservoir interval general inflow prediction system, which comprises a data standardization processing module, a short-term inflow prediction module, a medium and long-term inflow prediction module and a review analysis module. The short-term inflow prediction module obtains short-term inflow of each section; the data standardization processing module obtains surface rainfall and surface evaporation data; the medium and long-term inflow prediction module obtains medium and long-term inflow of each section; and the review analysis module analyzes and compares the prediction result saved during inflow prediction and the historical measured data saved by periodical processing. The multi-model fusion multi-reservoir interval general inflow prediction method and system can realize quick adaptation of short-term and medium and long-term inflow prediction of multiple reservoirs, intervals and sections, reduce the adaptation cost of model application, and improve the efficiency.

Description

Technical Field

[0001] This invention relates to the field of hydrological forecasting technology, and in particular to a multi-model fusion, multi-reservoir interval universal inflow forecasting system. Background Technology

[0002] In existing technologies, inflow forecasting systems generally use historical runoff data or future meteorological and rainfall data as input to predict reservoir inflow. The forecast results can be used to support flood control scheduling during the flood season and water resource allocation during the dry season, providing strong data support for the final implementation of reservoir scheduling plans. Inflow forecasts can be divided into two categories based on time scale: short-term inflow forecasts and medium- to long-term inflow forecasts.

[0003] Research on inflow forecasting systems began as early as 2018-2019 ("Research and Application of Ecological Dispatching of the Lijiang Reservoir Group", Fan Fuxin, 2018.6); however, the proposed solutions have certain limitations and the following problems urgently need to be addressed:

[0004] 1. Current research on historical inflow forecasts for the Lijiang River basin only considers medium- to long-term runoff forecasts. While medium- to long-term inflow forecasts provide an overall prediction of future inflow trends at the Guilin section, offering a general basis and direction for comprehensive water replenishment plans based on annual, monthly, and ten-day periods, they do not provide effective real-time water replenishment data for scenarios requiring water replenishment to allow cruise ship navigation due to insufficient water volume. For this scenario, short-term inflow forecasts can provide a basis for water replenishment at sections urgently needing replenishment in a short period. Short-term inflow forecasting for the Lijiang River basin is a previously unexplored topic.

[0005] 2. Due to the scarcity of original input data, the medium- and long-term inflow forecasting model for the Lijiang River basin only considers the influence of historical runoff sequences on the forecast sequence. Inflow forecasting models are heavily influenced by meteorological factors related to runoff (such as rainfall and temperature), and considering only the impact of historical runoff on future runoff is rather one-sided.

[0006] 3. Historical research on inflow forecasts for the Lijiang River basin only made medium- and long-term forecasts for the Guilin section. Detailed analysis and research were not conducted on inflow forecasts for the reservoirs upstream of Guilin Station (Qingshitan Reservoir, Xiaorongjiang Reservoir, Fuzikou Reservoir, and Chuanjiang Reservoir), the section between Guilin and Siku, and the section between Guilin and Yangshuo. The inflow forecasts only focused on the Guilin section, making it difficult to accurately replenish water in each section.

[0007] 4. Previous research on the inflow forecast of the Li River has been limited to the analysis, forecasting and verification of runoff based on historical data. It has failed to make real-time forecasts of future runoff based on meteorological factors, and to review and evaluate the accuracy of previous forecasts in real time. Past research has lacked practical value in actual application.

[0008] 5. During the reinforcement and upgrading of upstream reservoirs, the reservoirs cannot carry out normal regulation and storage. In this scenario, there is a lack of research and analysis on the inflow forecast for the cross-section. Summary of the Invention

[0009] The present invention aims to solve at least one of the technical problems mentioned above, and to provide a multi-model fusion, multi-reservoir interval universal water inflow forecasting system.

[0010] To achieve the above objectives, a multi-model fusion inflow forecasting system applicable to multiple reservoir areas is proposed, comprising a data standardization processing module, a short-term inflow forecasting module, a medium- and long-term inflow forecasting module, and a retrospective analysis module, wherein:

[0011] The data standardization processing module processes the hourly reported rainfall, evaporation, and flow data from the monitoring stations, and outputs the processing results to the short-term water inflow forecast module, the medium- and long-term water inflow forecast module, and the recap module. Data processing includes:

[0012] (1) Convert monitoring point data into area data: Design a station weight table to store the weights of the stations within the reservoir interval, find the corresponding weights, and then calculate the weighted sum of the rainfall and evaporation values ​​of the tuples in the list according to the weights to obtain hourly area rainfall and area evaporation data;

[0013] The input data for the medium- and long-term water inflow forecast module is processed by converting hourly areal rainfall data into ten-day and monthly average areal rainfall data. The calculation formula is as follows:

[0014] ,

[0015] in This indicates the average rainfall over a ten-day period / monthly period. The cumulative rainfall data for hour t reported in hour t, where N is the total number of hours in the ten-day period or month;

[0016] The formula for processing the ten-day / monthly average measured flow data used for medium- and long-term inflow forecasting is as follows:

[0017] ,

[0018] in This indicates the average rainfall over a ten-day period / monthly period. The instantaneous flow rate at hour t is expressed in m³ / s, and N is the total number of hours in that ten-day period or month.

[0019] (4) Design a database table structure to store the ten-day or monthly average flow data of each reservoir / section, and calculate the flow data in a timed manner through a timed task, and save the data to the medium and long-term inflow forecast data table;

[0020] The short-term inflow forecast module calculates the inflow of each reservoir based on the geographical location of the river and reservoir, using the output of the data standardization processing module as input. Through conventional reservoir scheduling algorithms or data fusion under the condition of risk elimination and reinforcement, the outflow of the reservoir is obtained, and the inflow of the interval is superimposed to obtain the inflow of each section.

[0021] The medium- and long-term inflow forecasting module takes forecasting factors as input, obtains forecast values ​​through BP neural network and K-nearest neighbor algorithm respectively, and then calculates the weighted average of the BP neural network forecast value and the K-nearest neighbor algorithm forecast value to forecast the medium- and long-term inflow of reservoirs and intervals. Based on the inflow of reservoirs and intervals, and whether the reinforcement and safety improvement scenario is configured, the outflow is calculated by combining the medium- and long-term conventional scheduling module and the reinforcement and safety improvement treatment unit. Finally, the inflow of the section is calculated by superimposing the inflow of the interval.

[0022] The retrospective analysis module analyzes and compares the forecast results saved during the water forecast and the historical measured data saved periodically, and evaluates and analyzes the correlation and error between the forecast flow series and the measured flow series.

[0023] Preferably, the control area of ​​each station is calculated using the Thiessen polygon method, and the weight of a single station is obtained by dividing the station's control area by the total control area of ​​all stations. Let the control area of ​​the polygon containing the i-th station be denoted as... The total controlled area is Then the weight of the station and satisfy Where n is the number of monitoring stations, and the areal rainfall is calculated as a weighted average of the rainfall monitored by each monitoring station in the area:

[0024] ,

[0025] Where m is the number of monitoring stations in the area, P is the areal rainfall, and p is the rainfall monitored by the monitoring station.

[0026] Preferably, the data standardization processing module includes a general data processing unit. The general data processing unit acquires the forecast rainfall sequence, evaporation sequence and initial flow reported by the radar station, calculates the areal rainfall and areal evaporation data within the control range of the reservoir or section, and inputs the areal data into the Xin'anjiang runoff model to calculate the forecast flow of the reservoir or section.

[0027] Preferably, the short-term and medium-to-long-term inflow forecast modules include a reinforcement and safety condition processing unit. This unit uses a hash table data structure to store three types of information for each reservoir: whether it is undergoing reinforcement and safety condition processing, the start and end times of the reinforcement and safety condition processing, and the discharge method. Based on the reservoir configuration, for reservoirs undergoing reinforcement and safety condition processing, if the discharge type is fixed discharge, the discharge flow rate is the flow rate at the moment before the reinforcement and safety condition processing begins; if the discharge type is natural runoff, the discharge flow rate is set to the inflow rate of the reservoir during that period.

[0028] Preferably, the short-term inflow forecast module includes a cross-sectional short-term inflow forecast calculation unit. The cross-sectional short-term inflow forecast calculation unit uses the inflow forecast results of upstream reservoirs at the cross-section. For reservoirs undergoing reinforcement, the unit inputs the forecast inflow of the reservoirs to the reinforcement working condition processing module. Otherwise, it uses a conventional reservoir scheduling algorithm to obtain the outflow of each reservoir and superimposes the inflow of the interval to finally obtain the cross-sectional flow.

[0029] Preferably, the reservoir discharge flow rate is calculated using the Muskingen model, and the calculation equation is as follows:

[0030] ,in, , .

[0031] Preferably, the BP neural network includes a model input layer, hidden layers, and an output layer. The input data is a time series of forecast factor vectors for the forecast period. For each vector, the data is first normalized, that is, the data of each dimension is normalized to between 0 and 1. Then, the normalized vector is input to the input layer of the neural network. Each feature element in the vector is processed by the neuron y=Wx+b, where W and b are the weight and bias, respectively. Then it is propagated to the neuron nodes of the hidden layer. After receiving the output of the neuron nodes of the input layer, the hidden layer neuron nodes are activated by the activation function y=1 / (1+exp(-x)) and propagated to the neuron nodes of the output layer in the same way. The output layer outputs the result through the same activation function. The output value of the neural network is denormalized to obtain the predicted runoff.

[0032] Preferably, the K-nearest neighbor algorithm calculates the Euclidean distance between the normalized forecast factor vector and the vectors in the historical dataset, sorts them by max-heap, and extracts the top K smallest vectors. For the top K smallest vectors, the weighted average of their flow is calculated according to the sorting results to obtain the forecast flow.

[0033] Preferably, the selection of the forecasting factors is based on the following criteria: the correlation coefficient is used to measure the correlation between the forecasting factors and runoff, and the correlation coefficient is calculated using the following formula:

[0034] ,

[0035] in, This represents the correlation coefficient between a forecasting factor and runoff q. This represents the multi-year average of a certain factor. , This represents the average sample runoff. , where l is the number of years in the runoff sample.

[0036] By calculating the correlation coefficients between each factor x and the runoff q, the correlation coefficient values ​​are arranged in descending order. Factors with the highest correlation coefficients are selected as initial factors. Then, the pairwise cross-correlation of the initial factors is analyzed according to the above formula, and factors with large correlation coefficients and small pairwise correlations are selected.

[0037] Preferably, the evaluation indicators of the review analysis module include the coefficient of certainty, absolute error, relative error, and pass rate, wherein:

[0038] The coefficient of determination mainly describes the degree of agreement between the predicted and measured flow processes, and its calculation formula is as follows:

[0039] (1.1),

[0040] in Let i be the predicted flow rate at time i. Let be the measured flow rate at time i. is the mean of the measured sequence, and n is the length of the predicted sequence;

[0041] The absolute error is the difference between the predicted value and the measured value for each time point. The formula for calculation is:

[0042] (1.2),

[0043] Relative error is the proportion of absolute error relative to the measured value, and its calculation formula is:

[0044] (1.3),

[0045] The pass rate is calculated by dividing the cumulative number of successful forecasts by the total number of forecast periods. The formula is as follows:

[0046] (1.4)

[0047] The isHege() function takes the measured sequence and the forecast sequence as input, calculates whether the forecast result at that time point is qualified according to different time scale requirements, and returns 1 if qualified, otherwise returns 0.

[0048] The beneficial effects are as follows: Compared with the existing technology, the multi-model fusion inflow forecasting system applicable to multiple reservoir sections of the present invention fills the gap in the short-term inflow forecasts of various reservoir sections in the Lijiang River by fusing real-time and forecast data; it improves and optimizes the structure of medium- and long-term forecast models to adapt to multi-dimensional data inputs (such as rainfall, temperature, historical runoff, etc.) to achieve more accurate and realistic forecast results; it supplements the inflow forecasts of various reservoirs and sections in the upper reaches of the Lijiang River, providing a basis for how to regulate water in each reservoir; it adds a historical review module to the historical forecast results, providing a reliable basis for the accuracy of the forecasts and assisting users in making more accurate water regulation decisions; for the inflow forecast of sections when the reservoir loses its regulation function under the condition of reinforcement and renovation, a configurable solution is proposed, which allows configuration of the start time of reinforcement and renovation, as well as the discharge mode of the reservoir during that period, and inflow forecasts of sections under this discharge condition. Attached Figure Description

[0049] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, wherein:

[0050] Figure 1 A schematic diagram of the general short-term inflow forecast module for a single reservoir or area;

[0051] Figure 2 This is a structural diagram of the Xin'anjiang model;

[0052] Figure 3 A schematic diagram of the processing flow for the hazard mitigation and reinforcement work unit;

[0053] Figure 4 This is a schematic diagram of the processing flow of the cross-sectional short-term inflow forecast calculation unit;

[0054] Figure 5 This is a flowchart illustrating the process of calculating inter-regional water inflow in the short-term water inflow forecast within the retrospective analysis module.

[0055] Figure 6 A schematic diagram of a general medium- to long-term inflow forecasting process for a single reservoir or watershed. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0057] It should be noted that when a component is described as "fixed to" another component, it can be directly on the other component or may have a central component. When a component is described as "connected to" another component, it can be directly connected to the other component or may have a central component. When a component is described as "set on" another component, it can be directly set on the other component or may have a central component. When a component is described as "set in the middle," it is not simply set in the exact center, as long as it is not set within the area defined by both ends being in the middle. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0058] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0059] This invention discloses a multi-model fusion, multi-reservoir interval universal inflow forecasting system, comprising a data standardization processing module, a short-term inflow forecasting module, a medium- and long-term inflow forecasting module, and a retrospective analysis module, wherein:

[0060] The data standardization processing module processes the hourly reported rainfall, evaporation, and flow data from the monitoring stations, and outputs the processing results to the short-term water inflow forecast module, the medium- and long-term water inflow forecast module, and the recap module. Its data processing includes:

[0061] (1) Convert monitoring point data into area data: Design a station weight table to store the weights of the stations within the reservoir interval, find the corresponding weights, and then calculate the weighted sum of the rainfall and evaporation values ​​of the tuples in the list according to the weights to obtain hourly area rainfall and area evaporation data;

[0062] The input data for the medium- and long-term water inflow forecast module is processed by converting hourly areal rainfall data into ten-day and monthly average areal rainfall data. The ten-day / monthly average rainfall is calculated by accumulating the hourly rainfall within the ten-day / monthly period and then dividing by the total number of hours in that period. The calculation formula is as follows:

[0063] ,

[0064] in This indicates the average rainfall over a ten-day period / monthly period. t represents the cumulative rainfall data reported in hour t, where N is the total number of hours in that ten-day period or month.

[0065] Processing of ten-day / monthly average measured flow data for medium- to long-term inflow forecasting and feedback. Based on the hourly flow data reported by flow stations, convert it into hourly runoff. Then, sum the runoff from all hours within the ten-day / month period, and divide by the number of seconds in that ten-day / month period to obtain the average flow for that ten-day / month period. The calculation formula is:

[0066] ,

[0067] in This indicates the average rainfall over a ten-day period / monthly period. The instantaneous flow rate at hour t is expressed in m³ / s, and N is the total number of hours in that ten-day period or month.

[0068] (4) Design the database table structure (medium and long term water inflow forecast and reconciliation data table) to store the ten-day or monthly average flow data of each reservoir / section, and calculate the flow data according to method (3) through a timed task (ten-day trigger) and save it to the medium and long term water inflow forecast and reconciliation data table.

[0069] The short-term inflow forecast module calculates the inflow of each reservoir based on the geographical location of the river and reservoir, using the output of the data standardization processing module as input. Through conventional reservoir scheduling algorithms or data fusion under the condition of risk elimination and reinforcement, the outflow of the reservoir is obtained, and the inflow of the interval is superimposed to obtain the inflow of each section.

[0070] The medium- and long-term inflow forecasting module takes forecasting factors as input, obtains forecast values ​​through BP neural network and K-nearest neighbor algorithm respectively, and then takes a weighted average of the BP neural network forecast value and the K-nearest neighbor algorithm forecast value to forecast the medium- and long-term inflow of reservoirs and intervals. Based on the inflow of reservoirs and intervals, combined with the medium- and long-term conventional scheduling model, the cross-sectional inflow is calculated.

[0071] The retrospective analysis module analyzes and compares the forecast results saved during the water forecast and the historical measured data saved periodically, and evaluates and analyzes the correlation and error between the forecast flow series and the measured flow series.

[0072] Specifically, this application uses the Lijiang River basin as one example for illustration. The short-term inflow forecast module first divides the reservoirs, sections, and cross-sections within the Lijiang River basin based on the geographical location of the river channel and reservoirs using the Thiessen polygon method. The division results are as follows: upstream reservoirs of Guilin Station: Qingshitan, Fuzikou, Chuanjiang, and Xiaorongjiang; sections include: the section from Guilin to Siku; the section from Guilin to Yangshuo is divided into: Guilin to Mopanshan, Mopanshan to Caoping, Caoping to Yangdi, Yangdi to Xingping, and Xingping to Yangshuo. The flow at Chaotian Station also flows towards Yangshuo through the river channel. The flow at Chaotian Station is calculated by superimposing the flow from the Si'anjiang Reservoir to the flow in the Si'anjiang-Chaotian section using the Muskingen model.

[0073] The short-term inflow forecast module first calculates the inflow of water to each reservoir upstream of the Guilin section. Then, through data fusion using conventional reservoir scheduling algorithms or under reinforcement and safety improvement conditions, it obtains the reservoir outflow. This outflow is then superimposed with the inflow from the surrounding area to obtain the inflow at each station (section). For example... Figure 1 As shown, the short-term inflow forecast module includes a general data processing unit for executing conventional reservoir scheduling algorithms. This unit acquires the forecast rainfall sequence, evaporation sequence, and initial flow rate reported by radar stations, calculates the area data within the reservoir or interval control area, and inputs this area data into the Xin'anjiang runoff generation model to calculate the forecast flow rate of the reservoir or interval. The Xin'anjiang model was proposed by Professor Zhao Renjun of Hohai University in 1963, and its modeling basis is the principle of runoff generation in humid regions. The Xin'anjiang model is a decentralized conceptual model that has been widely used in humid and semi-humid regions of my country. This scheme adopts the Xin'anjiang model with three water sources, its main characteristics being three divisions: unit division, water source division, and stage division. The watershed is divided into units, primarily to account for the uneven distribution of rainfall and secondarily to accommodate the different and varying underlying surface conditions. Water source differentiation involves dividing runoff into three components: surface, subsurface, and groundwater. These three sources have different confluence velocities, with surface runoff being the fastest and groundwater the slowest. Stage differentiation divides the runoff process into a slope confluence stage and a river network confluence stage. This is because the two stages have different confluence characteristics; on slopes, the confluence velocities of various water sources differ, while this difference is not observed in river networks. The model structure diagram is shown below. Figure 2 As shown, since each module of the Xin'anjiang model already has a specific implementation method, there are no major optimizations or modifications in this application module, so it will not be described again here.

[0074] like Figure 3 As shown, the short-term and medium-to-long-term inflow forecast modules also include a reinforcement and safety condition processing unit. This unit uses a hash table data structure to store three types of information for each reservoir: whether it is undergoing reinforcement and safety condition processing, the start and end times of the reinforcement and safety condition processing, and the discharge method. Based on the reservoir configuration, for reservoirs undergoing reinforcement and safety condition processing, if the discharge type is fixed discharge, the discharge flow rate is the flow rate at the moment before the reinforcement and safety condition processing begins; if the discharge type is natural runoff, the discharge flow rate is set to the inflow rate of the reservoir during that period.

[0075] In addition, such as Figure 4 As shown, the short-term inflow forecast module includes a cross-sectional short-term inflow forecast calculation unit. This unit uses forecasts from the upstream reservoir inlet, the Yangtze River, and the interval between the reservoir and the cross-section. For reservoirs undergoing reinforcement, the forecasted inflow is input to the reinforcement condition processing module; otherwise, a conventional reservoir scheduling algorithm is used to obtain the outflow from each reservoir, which is then superimposed with the interval inflow to finally obtain the cross-sectional flow. Specifically, the reservoir outflow is calculated using the Muskingen model, and its calculation equation is as follows:

[0076] ,in In the above formula, the flow rate at the lower section during the second time period is equal to the flow rate at the upper section during the second time period multiplied by a coefficient. Add the first period's cross-sectional flow rate multiplied by Add the cross-sectional flow rate of the first time period multiplied by a coefficient. Where K represents the time it takes for the water flow to travel from the upper cross section to the lower cross section, and x is the flow rate density coefficient. The calculation period is defined as follows. For a continuous flow sequence, based on the above formula, the evolved flow of the lower section can be calculated from the flow of the upper section. The Muskingan module mainly calculates the result of the evolution of the flow of the upper section to the flow of the lower section at each moment according to the above formula. Its purpose is to ensure that at the same moment, the composition of the flow of this section is the result of the evolution of the upstream flow.

[0077] In this embodiment, by using the data standardization processing module for scenarios involving multiple monitoring stations within a reservoir or interval, the system can directly associate the reservoir (or interval) to which a monitoring station belongs and configure its weight. This allows data from that station to be used as input data for inflow forecasting without requiring a restart or modification of the application, thus improving the scalability of monitoring station data access. The data standardization processing module primarily processes the monitoring point data reported by each reservoir or interval, converting it into area data. Rainfall and evaporation data for the reservoir or interval control area are obtained by multiplying the measured values ​​of each monitoring point within the control area by their respective weights, ensuring that the sum of the weights of all monitoring stations within the basin is 1. The data standardization processing module employs a station weight table to store the weights corresponding to the monitoring stations within their respective reservoir or interval ranges. For a reservoir or region, the data standardization processing module receives a sequence of tuples, i.e., the data reported by the monitoring stations. The tuple attributes are <station code, evaporation, rainfall, time>. The processing logic first groups the data by time, creating a list of tuples from the same time period. Then, based on the station value of each tuple in the list, its corresponding weight is found. Finally, the rainfall and evaporation values ​​of the tuples in the list are weighted and summed to obtain the areal rainfall and areal evaporation data. This process is repeated for each time period's list to calculate the areal rainfall and evaporation data. Specifically, the station weight table is shown below:

[0078]

[0079] The control area of ​​each rain gauge station within the watershed is calculated using the Thiessen polygon method. In the Thiessen polygon method, all adjacent rain gauge stations are connected to form triangles. The perpendicular bisectors of each side of these triangles are then drawn, and the polygons enclosed by these perpendicular bisectors around each rain gauge station are the Thiessen polygons. The weight of a single station is obtained by dividing its control area by the total control area of ​​all stations. Let the control area of ​​the polygon containing the i-th station be denoted as . The total controlled area is Then the weight of the station , and satisfy Where n is the number of monitoring stations. The areal rainfall is calculated as a weighted average of the rainfall monitored by all stations in the region:

[0080] ,

[0081] Where P is the areal rainfall and p is the rainfall monitored at the station.

[0082] In this embodiment, Figure 6 The diagram illustrates the calculation process of the medium- and long-term inflow forecasting module. This module takes forecasting factors as input and obtains forecast values ​​through both a backpropagation (BP) neural network and a K-nearest neighbor (KNN) algorithm. The BP neural network and KNN forecast values ​​are then weighted and averaged to predict the medium- and long-term inflows of the reservoir and its intervals. Based on the inflows of the reservoir and its intervals, and considering whether the reservoir is undergoing reinforcement and upgrading as configured by the user, the outputs of the reinforcement and upgrading module and the medium- and long-term routine scheduling module are combined with the interval inflow to calculate the cross-sectional inflow. The medium- and long-term routine scheduling module determines the reservoir's discharge volume based on the reservoir's scheduling map, river irrigation water demand, and water replenishment requirements.

[0083] The selection of forecast factors is based on the following criteria: the correlation coefficient is used to measure the correlation between the forecast factors and runoff. The formula for calculating the correlation coefficient is as follows:

[0084] ,

[0085] in, This represents the correlation coefficient between a forecasting factor and runoff q. This represents the multi-year average of a certain factor. , This represents the average sample runoff. , where l is the number of years in the runoff sample.

[0086] By calculating the correlation coefficients between each factor x and runoff q, the correlation coefficient values ​​are arranged in descending order. Factors with the highest correlation coefficients are selected as initial factors. Then, the cross-correlation of the initial factors is analyzed pairwise according to the above formula, and factors with large correlation coefficients and low pairwise correlations are selected. Finally, through calculation, rainfall, temperature, and upper time period runoff are selected as the forecasting factors, and the forecasting factors are input into the BP neural network and K-nearest neighbor algorithm.

[0087] The BP neural network consists of an input layer, hidden layers, and an output layer. The input data is a vector time series of forecast factors (rainfall, temperature, and runoff in the previous period) for the forecast period. For each vector, the data is first normalized, that is, the data of each dimension is normalized to the range of 0-1. Then, the normalized vector is input into the input layer of the neural network. Each feature element in the vector is processed by the neuron y=Wx+b, where W and b are the weight and bias, respectively. Then it is propagated to the neurons in the hidden layer, and each hidden layer neuron receives the output of the neurons in the input layer. Then, through the activation function y=1 / (1+exp(-x)), the propagation is carried out to the nodes of the output layer neurons in the same way. The output layer outputs the result through the same activation function. The neural network output value is denormalized to obtain the predicted runoff. The BP neural network in this embodiment considers the factors affecting runoff, that is, it is believed that rainfall, temperature, and runoff in the previous period have a significant impact on the runoff in the next period, and these features are independent of each other and there is no dependency. Compared with previous studies that only considered the single factor of historical runoff characteristics, the factors considered in this embodiment are more detailed. The training algorithm of the neural network uses the back diffusion algorithm to initialize the parameters of each neuron, that is, W,b (weights and biases), to train the BP neural network. In order to avoid overfitting of the BP neural network, the input used to calculate the error uses the predicted value and the measured value of the validation set, respectively. When the error is less than a certain threshold, the training stops and the BP neural network parameters are saved.

[0088] The K-Nearest Neighbors (KNN) algorithm calculates the Euclidean distance between the normalized forecast factor (rainfall, temperature, and previous period runoff) vectors and the vectors in the historical dataset (the number of vectors is optional, i.e., the value of K). It then sorts these vectors using a max-heap and extracts the top K smallest vectors. For each of these top K smallest vectors, a weighted average of their flow rates is calculated based on the sorting results and weights (higher-ranked vectors have higher weights, with all weights summing to 1). This weighted average yields the forecast flow rate. Specifically, the formula for calculating the forecast flow rate using the KNN algorithm is as follows: Assuming the top K data items with the highest Euclidean distance from the historical dataset have been identified using the feature factor vectors of the flow rate to be forecasted, their corresponding flows are... Then the flow rate to be predicted The calculation formula is:

[0089] (2.1),

[0090] in The calculation method used was the method suggested in the literature (Upmanu Lall, Ashish Sharma. A nearest neighbor bootstrap for resampling hydrologic time series. Water Resource Research, 1996), that is...

[0091] (2.2),

[0092] Where j = 1, 2, ..., k. The weighting formula shows that among the top K data items, the smaller the Euclidean distance between their feature factors and the Euclidean distance between the feature factors of the flow to be predicted, the higher the weight they are assigned when calculating the flow. The following example demonstrates using the K-nearest neighbor algorithm to predict flow. Suppose the predicted flow for a certain ten-day period is... Its feature factor vector is assumed to be (Average rainfall = 3.0, average temperature = 35.2), assuming the historical dataset contains the following data: x1 (average rainfall = 3.2, average temperature = 36.0, flow rate = 23), x2 (average rainfall = 3.5, average temperature = 34.0, flow rate = 28), x3 (average rainfall = 0.5, average temperature = 33.0, flow rate = 18). By normalizing the feature vectors, the result is the normalized feature factor vector of the data to be predicted, which is ( =0.83, Similarly, the eigenvectors of x1 to x3 can be obtained as (0.9,1), (1,0.33), and (0,0), respectively. Assuming the value of k is 2, after normalization... The Euclidean distances between x1 and x3 can be used to determine the relationship between x1 and x3. The closest top 2 vectors are x1 and x2, respectively. According to equation (2.2), we obtain... , According to equation (2.1), we can obtain... .

[0093] In this embodiment, the medium- and long-term water inflow forecast module uses a weighted average of the forecast values ​​from a BP neural network and the K-nearest neighbor algorithm to obtain the final result. The reason for combining these two methods is that the BP neural network has some uncertainty in forecasting extreme data (such as extreme rainfall or temperature), but has high accuracy for normal inputs. The K-nearest neighbor algorithm is a deterministic algorithm; given a fixed historical dataset, its calculated result will not change. Combining these two algorithms ensures that the medium- and long-term forecast system has good forecast accuracy under normal conditions, while also maintaining forecast stability under extreme input data conditions.

[0094] This embodiment also includes a retrospective analysis module. This module analyzes and compares the forecast results saved during water inflow forecasting (including short-term and medium-to-long-term water inflow forecasts) with periodically processed and saved historical measured data. This allows for an understanding of the accuracy of the water inflow forecasts and timely adjustments. Taking the Lijiang River basin water inflow forecast retrospective as an example, the retrospective analysis module includes a water inflow forecast result table and a water inflow forecast retrospective data table. The water inflow forecast result table is shown below:

[0095]

[0096] The table below shows the expected return data:

[0097]

[0098] As shown in the table above, the forecast data from the review and analysis module is centrally stored in the forecast results table. For each forecast, the following data is stored: a unique forecast identifier (call_id), forecast type (predict_type) (marking the forecast as a short-term or medium-to-long-term inflow forecast), forecast flow (predict_flow) and time (time) for each time period, and the reservoir / section / cross-section number corresponding to that forecast. The time interval for each data point in the forecast varies depending on the forecast type. For example, the time interval for short-term inflow forecast data is 1 hour, and the time interval for medium-to-long-term inflow forecast data is 1 ten-day period (e.g., the first ten-day period is represented by xxxx year x month 1st, the middle ten-day period by xxxx year x month 11th, and the last ten-day period by xxxx year x month 21st). For the acquisition of retrospective data, since the time interval for data reported by each flow monitoring station is hourly, an additional timer (triggered every ten days) is designed to call the monitoring data service every ten days to obtain hourly flow data within a certain time range. The retrospective measured (backward) flow data is obtained from the inflow forecast retrospective data table. For medium- and long-term inflow forecasts, the average measured flow data is completed in the data standardization processing module and updated to the forecast retrospective data table periodically through a scheduled task; short-term inflow forecasts are updated into this table based on the station reports. After data cleaning (smoothing outlier data, filling in missing data, etc.), the average flow for each ten days or month is calculated, and the average value is saved to the medium- and long-term inflow forecast retrospective data table according to the station type. The flow types include the instantaneous flow (type is hourly data) of reservoir outflow and inflow, the average flow (type is ten days or monthly data), and the flow through the cross section. The calculation methods for the indicators in the retrospective analysis are given in formulas (1.1~1.4). The qualification standards are based on the short-term, medium-term, and long-term qualitative and quantitative accuracy assessments defined in the "GBT22482-2008 Hydrological Information Forecasting Specification." Specific parameter values ​​are not repeated here; only the source and storage format of the data to be analyzed are described. The data table used in the retrospective analysis module stores intermediate measured data. When using the retrospective function, it avoids directly requesting monitoring and processing from monitoring stations; instead, it directly reads data from the database table for assessment, improving retrospective efficiency. The data standardization processing module uses a background timer to periodically trigger, process, and save monitoring data, effectively supporting the analytical capabilities and performance of the retrospective analysis module.

[0099] Specifically, the retrospective analysis module mainly analyzes the correlation and error between the forecasted and measured flow sequences. The inputs to this module are the forecasted and measured flow sequences (or inversely derived flow sequences), and its evaluation indicators include the coefficient of certainty, absolute error, relative error, and pass rate.

[0100] The coefficient of determination mainly describes the degree of agreement between the predicted and measured flow processes, and its calculation formula is as follows:

[0101] ,

[0102] in Let i be the predicted flow rate at time i. Let be the measured flow rate at time i. is the mean of the measured sequence, and n is the length of the predicted sequence;

[0103] The absolute error is the difference between the predicted value and the measured value for each time point. The formula for calculation is as follows:

[0104] ,

[0105] Relative error is the proportion of absolute error relative to the measured value, and its calculation formula is:

[0106] ,

[0107] The pass rate is calculated by dividing the cumulative number of successful forecasts by the total number of forecast periods. The formula is as follows:

[0108] ,

[0109] The isHege() function takes the measured sequence and the forecast sequence as input, calculates whether the forecast result at that time point is qualified according to different time scale requirements, and returns 1 if qualified, otherwise returns 0.

[0110] In the retrospective analysis module, since there are no measured flows for the interval, this module proposes a method to calculate the inflow for the interval in the short-term inflow forecast by using inverse calculation of the evolved flow rate, in order to evaluate the accuracy of the short-term inflow forecast for the interval. The specific calculation process is as follows: Figure 5 As shown. Since water flow requires evolution time, the flow rate at the same time interval cannot be calculated by subtracting the reservoir outflow from the flow rate at the same cross-section at the same time. Therefore, this calculation method considers the evolution process after the reservoir outflow, and uses the Muskingen evolution equation to calculate the flow rate of each reservoir outflow reaching the cross-section at each time point.

[0111] As can be seen from the above embodiments, the water inflow forecasting system of the present invention has the following advantages compared with the prior art:

[0112] 1. This scheme fills the gap in historical research on short-term inflow forecasts for various reservoirs, sections, and cross sections of the Lijiang River. The relevant results of short-term inflow forecasts have been successfully applied to the "Digital Twin Lijiang River Water Resources Allocation 'Four Forecasts'" system. The short-term inflow forecast results for reservoirs and sections are good, and the average coefficient of certainty reaches about 0.8 based on the evaluation results of historical flood events.

[0113] 2. This scheme improves and optimizes the structure of the medium- and long-term inflow forecasting model based on the Lijiang River Basin. For the first time, it proposes a model adaptable to multi-dimensional data inputs (rainfall, temperature, historical runoff, etc.) in the Lijiang River Basin. Furthermore, by optimizing the parameter calibration strategy of the BP neural network, the model achieves a better fit during training with historical data. This achievement has been successfully applied to the "Digital Twin Lijiang River Water Resources Allocation 'Four Forecasts'" system. Review results show that the qualitative and quantitative forecasting qualification rates are above 70%.

[0114] 3. This solution proposes a standardized data processing module. For scenarios with multiple monitoring stations within a reservoir or interval, the reservoir (or interval) to which the monitoring station belongs and its weight can be directly configured in the database table. The data from that monitoring station can then be used as input data for water inflow forecasting without restarting or modifying the application, thus improving the scalability of monitoring station data access.

[0115] 4. This plan reasonably processes forecast data and historical measured data, and develops a retrospective analysis module to evaluate the accuracy of water inflow forecasts for various reservoirs, sections, and cross-sections within the Lijiang River basin.

[0116] 5. This solution provides a configuration method for the discharge during the reservoir reinforcement and renovation period. Through coordination and invocation with various modules, it enables the forecasting of water inflow at the cross-section during the reservoir reinforcement and renovation period.

[0117] All of the above achievements have been applied to the "Digital Twin Lijiang River Water Resources Allocation 'Four-Pre-Control' System".

[0118] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of the technical solutions of the present invention.

Claims

1. A multi-model fusion, multi-reservoir interval universal inflow forecasting system, characterized in that, It includes a data standardization and processing module, a short-term water inflow forecasting module, a medium- and long-term water inflow forecasting module, and a retrospective analysis module, among which: The data standardization processing module processes the hourly reported rainfall, evaporation, and flow data from the monitoring stations, and outputs the processing results to the short-term water inflow forecast module, the medium- and long-term water inflow forecast module, and the recap module. Data processing includes: (1) Convert monitoring point data into area data: Design a station weight table to store the weights of the stations within the reservoir interval, find the corresponding weights, and then calculate the weighted sum of the rainfall and evaporation values ​​of the tuples in the list according to the weights to obtain hourly area rainfall and area evaporation data; (2) Process the input data of the medium- and long-term water inflow forecast module, and convert the hourly areal rainfall data into ten-day and monthly average areal rainfall data. The calculation formula is as follows: ， in This indicates the average rainfall over a ten-day period / monthly period. The cumulative rainfall data for hour t reported in hour t, where N is the total number of hours in the ten-day period or month; (3) The calculation formula for processing the ten-day / monthly average measured flow data used for medium- and long-term inflow forecasting is as follows: ， in This indicates the average rainfall over a ten-day period / monthly period. The instantaneous flow rate at hour t is expressed in m³ / s, and N is the total number of hours in that ten-day period or month. (4) Design a database table structure to store the ten-day or monthly average flow data of each reservoir / section, and calculate the flow data in a timed manner through a timed task, and save the data to the medium and long-term inflow forecast data table; The short-term inflow forecast module calculates the inflow of each reservoir based on the geographical location of the river and reservoir, using the output of the data standardization processing module as input. Through conventional reservoir scheduling algorithms or data fusion under the condition of risk elimination and reinforcement, the outflow of the reservoir is obtained, and the inflow of the interval is superimposed to obtain the inflow of each section. The medium- and long-term inflow forecasting module takes forecasting factors as input, obtains forecast values ​​through BP neural network and K-nearest neighbor algorithm respectively, and then takes a weighted average of the BP neural network forecast value and the K-nearest neighbor algorithm forecast value to forecast the medium- and long-term inflow of reservoirs and intervals. Based on the inflow of reservoirs and intervals, combined with the medium- and long-term conventional scheduling model, the cross-sectional inflow is calculated. The retrospective analysis module analyzes and compares the forecast results saved during the water forecast and the historical measured data saved periodically, and evaluates and analyzes the correlation and error between the forecast flow series and the measured flow series.

2. The multi-model fusion, multi-reservoir interval universal inflow forecasting system according to claim 1, characterized in that, The control area of ​​each station is calculated using the Thiessen polygon method, and the weight of a single station is obtained by dividing its control area by the total control area of ​​all stations. Let the control area of ​​the polygon containing the i-th station be denoted as . The total controlled area is Then the weight of the station and satisfy Where n is the number of monitoring stations within the total control area A, and the areal rainfall within the control area is calculated as a weighted average of the rainfall monitored by each monitoring station: ， Where P is the areal rainfall and p is the rainfall monitored at the station.

3. The multi-model fusion, multi-reservoir interval universal inflow forecasting system according to claim 1, characterized in that, The data standardization processing module includes a general data processing unit. The general data processing unit acquires the forecast rainfall sequence, evaporation sequence and initial flow reported by the radar station, calculates the areal rainfall and areal evaporation data within the control range of the reservoir or section, and inputs the areal data into the Xin'anjiang runoff model to calculate the forecast flow of the reservoir or section.

4. The multi-model fusion, multi-reservoir interval universal inflow forecasting system according to claim 1, characterized in that, The short-term and medium-to-long-term inflow forecast modules include a reinforcement and safety condition processing unit. This unit uses a hash table data structure to store three types of information for each reservoir: whether it is undergoing reinforcement and safety condition processing, the start and end times of the reinforcement and safety condition processing, and the discharge method. Based on the reservoir configuration, for reservoirs undergoing reinforcement and safety condition processing, if the discharge type is fixed discharge, the discharge flow rate is the flow rate at the moment before the reinforcement and safety condition processing begins; if the discharge type is natural runoff, the discharge flow rate is set to the inflow rate of the reservoir at the current moment.

5. The multi-model fusion, multi-reservoir interval universal inflow forecasting system according to claim 4, characterized in that, The short-term inflow forecast module includes a cross-sectional short-term inflow forecast calculation unit. The cross-sectional short-term inflow forecast calculation unit uses the inflow forecast results of upstream reservoirs at the cross-section. For reservoirs undergoing reinforcement, the unit inputs the forecast inflow of the reservoirs into the reinforcement working condition processing module. Otherwise, it uses a conventional reservoir scheduling algorithm to obtain the outflow of each reservoir and superimposes the inflow of the interval to finally obtain the cross-sectional flow.

6. The multi-model fusion, multi-reservoir interval universal inflow forecasting system according to claim 1, characterized in that, The BP neural network includes an input layer, hidden layers, and an output layer. The input data is a time series of forecast factor vectors for the forecast period. For each vector, the data is first normalized, that is, the data of each dimension is normalized to between 0 and 1. Then, the normalized vector is input into the input layer of the neural network, and each feature element in the vector is processed by a neuron. The operation is performed, where W and b are the weights and biases of the neuron, respectively. The feature vector is input to the neural network; The propagation reaches the hidden layer neurons, and each hidden layer neuron node... As input, through the activation function ,in The output of the input layer; The same method is used to propagate to the neuron nodes in the output layer. The output layer outputs the result through the same activation function. The neural network output value is then denormalized to obtain the predicted runoff.

7. The multi-model fusion inflow forecasting system applicable to multiple reservoir areas according to claim 1, characterized in that, The K-nearest neighbor algorithm calculates the Euclidean distance between the normalized forecast factor vector and the vectors in the historical dataset, sorts them by max-heap, and extracts the top K smallest vectors. For the top K smallest vectors, the weighted average of their flow is calculated according to the sorting results to obtain the forecast flow.

8. The multi-model fusion, multi-reservoir interval universal inflow forecasting system according to claim 1, characterized in that, The selection of the forecasting factors is based on the following criteria: the correlation coefficient is used to measure the correlation between the forecasting factors and runoff, and the formula for calculating the correlation coefficient is as follows: ， in, This represents the correlation coefficient between a forecasting factor and runoff q. This represents the multi-year average of a certain factor. , This represents the average sample runoff. , where l is the number of years in the runoff sample. By calculating the correlation coefficients between each factor x and the runoff q, the correlation coefficient values ​​are arranged in descending order. Factors with the highest correlation coefficients are selected as initial factors. Then, the pairwise cross-correlation of the initial factors is analyzed according to the above formula, and factors with large correlation coefficients and small pairwise correlations are selected.

9. The multi-model fusion, multi-reservoir interval universal inflow forecasting system according to claim 1, characterized in that, The evaluation indicators of the debriefing analysis module include the coefficient of certainty, absolute error, relative error, and pass rate, among which: The coefficient of determination is used to describe the degree of agreement between the predicted and measured flow processes, and its calculation formula is as follows: ， in Let i be the predicted flow rate at time i. Let be the measured flow rate at time i. is the mean of the measured sequence, and n is the length of the predicted sequence; The absolute error is the difference between the predicted value and the measured value for each time point. The formula for calculation is as follows: ， Relative error is the proportion of absolute error relative to the measured value, and its calculation formula is: ， The pass rate is calculated by dividing the cumulative number of successful forecasts by the total number of forecast periods. The formula is as follows: ， The isHege() function takes the measured sequence and the forecast sequence as input, calculates whether the forecast result at that time point is qualified according to different time scale requirements, and returns 1 if qualified, otherwise returns 0.

Images