TWI587222B - Neural networks based water level predicting system and method for reservoirs - Google Patents

Description

Reservoir water level prediction system and method based on neural network

The invention relates to a reservoir-based water level prediction system and method based on a neural network, in particular to a reservoir-based water level prediction system and method using a plurality of types of real-time hydrological data.

Hydropower plants must consider both flood control and water supply when considering how to generate electricity efficiently. Therefore, the power generation operation of hydropower plants is closely related to the water level of the reservoir. In the power generation operation, the current reservoir water level and the future reservoir water level must be taken into consideration together, otherwise it will not be able to achieve high power generation efficiency, flood control and water supply at the same time.

In addition, the amount of electricity generated by hydropower plants is also affected by weather and seasonal water volume. With the global climate change intensifying, Taiwan's subtropical Taiwan has become more and more frequent in recent years. In the hydropower plant power generation operation and reservoir water level prediction, various factors such as the rainfall observation value of the catchment area, the rainfall forecast value of the catchment area, and the flood discharge amount should be taken into consideration at the same time, so as to effectively predict the water level of the reservoir and thereby carry out the power generation operation. In order to achieve high power generation efficiency, flood control and water supply.

However, it is necessary to take such diverse and complex information into consideration. The water level of the reservoir is not achieved by manpower. At present, the power generation operation of hydropower plants is only carried out by the operating personnel of hydropower plants based on the current reservoir water level height value and some related information, based on personal experience judgment, and cannot accurately predict the reservoir water level. Therefore, there is a need for a reservoir water level prediction system and method that can be used to consider multiple factors related to reservoir water levels.

In order to solve the above problems, an object of the present invention is to provide a reservoir-based water level prediction system and method based on a neural network.

Another object of the present invention is to provide a neural network-based reservoir water level prediction system and method that can simultaneously incorporate various factors into prediction considerations.

A further object of the present invention is to provide a neural network-based reservoir water level prediction system and method for learning the relationship between reservoir water level and input data based on historical data and thereby predicting reservoir water level.

To achieve the above objective, the present invention provides a reservoir-based water level prediction system based on a neural network, comprising: a prediction module, a database, and a user operation interface. The prediction module further comprises: an input data storage module, a neural network learning module and a predictive data computing module. The input data storage module captures and stores at least one input data from a plurality of servers, and the input data includes at least real-time data and historical data of the reservoir water level. The neural network learning module molds at least one weight and at least one bias value according to the historical data stored in the input data storage module. The predictive data operation module calculates the current data according to the weight and the bias value molded by the neural network learning module to obtain at least one predicted data, and the predicted data includes at least a reservoir Forecast data for water level. And the database receives And storing the forecast data and/or the input data. The user operation interface receives at least one user operating condition and displays the predicted data and/or the input data and/or the predicted data of the predicted data according to the user operating conditions.

In a preferred embodiment of the present invention, the neural network learning module has a first layer, a second layer, and a third layer neural network.

In a preferred embodiment of the present invention, the neural network learning module molds the weight of the first layer neural network and the bias value according to at least the historical data of the water level of the reservoir, and the prediction data operation mode The group calculates the inflow data of the reservoir based on the weight of the first layer of the mold and the bias value, and the inflow data of the reservoir is the input data of the second layer neural network. One.

In a preferred embodiment of the present invention, the neural network learning module molds the weight of the second layer neural network and the bias value according to at least the inflow data of the reservoir, and the prediction data operation module Based on the weight of the second layer molding and the bias value, the forecast data of the inflow of the reservoir is calculated for at least the real-time output data of the first layer neural network, and the forecast data of the inflow of the reservoir is the third One of the input data of the layered neural network.

In a preferred embodiment of the present invention, the neural network learning module molds the weight of the third layer neural network and the bias value according to at least the predicted data of the reservoir inflow, and the prediction data operation The module calculates the predicted water level of the reservoir based on the weight of the third layer molding and the bias value to calculate the instantaneous output data of at least the second layer neural network.

In a preferred embodiment of the present invention, the neural network learning module molds the weight of the first-level neural network and the bias value according to historical data of power generation and flood discharge. The predictive data calculation module calculates the inflow data of the reservoir based on the weight of the first layer of the mold and the bias value to calculate at least the real-time data of the water level, the power generation amount and the flood discharge amount of the reservoir.

In a preferred embodiment of the present invention, the neural network learning module molds the weight and the bias of the second layer neural network according to the historical data of the rainfall station observation of the catchment area and the meteorological forecast of the catchment area. Weight, the predictive data operation module is based on the weight of the second layer of the mold and the bias value for at least the first layer of neural network real-time output data, the catchment rainfall station observation and the catchment area meteorological The forecasted real-time data is used to calculate the forecast data of the reservoir inflow.

In a preferred embodiment of the present invention, the neural network learning module molds the weight of the third-level neural network and the bias value according to historical data of power generation and flood discharge, and the prediction data operation The module calculates the forecast data of the water level of the reservoir based on the weight of the third layer molding and the bias value to calculate at least the real-time data of the instantaneous output data, the power generation amount and the flood discharge amount of the second layer neural network.

In a preferred embodiment of the invention, the predictive data includes reservoir inflow prediction data.

The input data includes real-time data and historical data of one or part of the power generation data, the flood discharge data, the rainfall station observation data of the catchment area, and the meteorological forecast data of the catchment area.

According to the object of the present invention, a reservoir-based water level prediction method based on a neural network is provided, which comprises: constructing a neural network model; determining historical data for training; and modeling a neural network based on the training historical data. The weight and the bias value; collecting at least one instant input data, the real-time input data includes at least the real-time data of the reservoir water level; and the input data according to the weight and the bias value molded by the neural network Perform an operation to get At least one forecast data, and the forecast data includes at least reservoir water level prediction data.

In a preferred embodiment of the invention, the neural network model has a plurality of layers of neural networks.

In a preferred embodiment of the present invention, at least one layer of neural network computing output data is reservoir inflow prediction data, and the reservoir inflow prediction data is input of a layer of neural network under the layered neural network. One of the materials.

In a preferred embodiment of the invention, the predictive data includes reservoir inflow prediction data.

In a preferred embodiment of the present invention, the instant input data includes real-time data of one or a part of power generation data, flood discharge data, rainfall station observation data of the catchment area, and meteorological forecast data of the catchment area.

The foregoing aspects and other aspects of the invention will be apparent from the description of the appended claims appended claims

100‧‧‧ class neural network system

200‧‧‧ Training materials

220‧‧‧Data Server

240‧‧‧ Reservoir Water Level Prediction System

250‧‧‧ Prediction Module

252‧‧‧ Neural Network Learning Module

254‧‧‧ Forecast Data Calculation Module

256‧‧‧Input data storage module

260‧‧‧User interface

262‧‧‧Operating module

264‧‧‧ Forecast data display module

270‧‧‧Database

280‧‧‧Users

300‧‧‧ forecasting mechanism flow chart

310-350‧‧‧ operation

400‧‧‧ class neural network model

402‧‧‧Power generation data

404‧‧‧ Reservoir water level data

406‧‧‧ Flood discharge data

408a‧‧‧First-class neural network

408b‧‧‧Second-level neural network

408c‧‧‧Layer 3 neural network

412‧‧‧ Reservoir Inflow Information

424‧‧‧Catchment observation data of catchment area

426‧‧‧Catchment meteorological forecast data

432‧‧‧ Reservoir Inflow Forecast Data

442‧‧‧ Reservoir water level forecast data

500‧‧‧User operation display flow chart

510-550‧‧‧ operation

610‧‧‧Graphic display area

612‧‧‧ vertical axis

614‧‧‧ Reservoir Inflow Forecasting Line Chart

616‧‧‧ Reservoir water level prediction line chart

618‧‧‧ vertical axis

619‧‧‧ horizontal axis

620‧‧‧Graphic display area

622‧‧‧ vertical axis

624‧‧‧ horizontal axis

626‧‧‧Watershed rainfall forecast bar chart

The first picture is a schematic diagram of a general neural network system.

The second figure is an architectural diagram of a reservoir-based water level prediction system based on a neural network.

The third figure is a flow chart of the method for predicting the water level of the reservoir based on the neural network.

The fourth figure is a neural network model diagram of a specific embodiment of a reservoir-based water level prediction system based on a neural network.

The fifth figure is a flow chart of the user operation display of the reservoir water level prediction method based on the neural network.

The sixth figure is a schematic diagram of a specific embodiment of the prediction result display of the reservoir-based water level prediction system based on the neural network.

The invention is based on a neural network-based reservoir water level prediction system and method thereof, which utilizes a neural network having the characteristics of learning the relationship between input data and output data under the condition that no mathematical transformation is required. The neural network is trained with a large amount of training data to mold the weights of each input node and each hidden layer neuron, the weight of each hidden layer neuron and the output node, and the partial weight of each node. The hydrological data obtained is used as input data for each input node to predict the reservoir water level. The following will be further explained in conjunction with the drawings.

The first picture is a schematic diagram of a general neural network system. The invention is based on a neural network-based reservoir water level prediction system, which is trained by a neural network system 100 as shown in the figure to mold various weights and respective partial weight values. In the figure, the neural network of the type The road system 100 has an input node S1, an input node S2 in the first layer, a hidden layer neuron S3, a hidden layer neuron S4, a hidden layer neuron S5 in the second layer, and an output node S6 in the third layer.

It should be understood that the neural network-like system 100 is merely exemplary herein, and that the present invention is not limited to the use of a single layer of hidden layer neurons, but one or more layers of hidden layer neurons may be used as desired. The number of input nodes, the number of hidden layer neurons, and the number of output nodes of the neural network system 100 are also not limited to the present invention. The number of input nodes, the number of hidden layer neurons, and the output section of the present invention are not limited. The number of points can be adjusted to any number depending on the situation.

Referring to the first figure, each input node of the neural network system 100 has a corresponding weight with each hidden layer neuron. For example, the input node S1 and the hidden layer neuron S5 have a weight W15, and the input node S2 is hidden. The layer neuron S5 has a weight W25. At the same time, each hidden layer neuron of the neural network system 100 also has a corresponding weight. For example, the hidden layer neuron S3 and the output node S6 have a weight W36, and the hidden layer neuron S5 and the output node S6 have a weight. Weight W56. Each input node, hidden layer neuron, and output node each have a respective bias value. For example, the input node S2 has an offset weight θ2, the hidden layer neuron S5 has an offset weight θ5, and the output node S6 has an offset weight θ6. The calculation of the transmission value transmitted by one node to the next node is as follows. Assuming that a total of n nodes transmit their respective transmission values to the node Y, the formula of the transmission value y transmitted by the node Y to the next node is: Where Wi is the weight corresponding to the i-th node and the node Y of the n nodes transmitting the transmission value to the node Y, Xi is the transmission value of the i-th node transmitted to the node Y, and θ is the deviation of the node Y Weight.

Taking the node S5 as an example, the weight corresponding to the input node S1 and the hidden layer neuron S5 is W15, the weight corresponding to the input node S2 and the hidden layer neuron S5 is W25, and the bias value of the hidden layer neuron S5 is θ5. . Let the transmission value of the input node S1 transmitted to the hidden layer neuron S5 be X1, and the transmission value of the input node S2 transmitted to the hidden layer neuron S5 is X2, then the transmission value of the hidden layer neuron S5 to the output node S6 is: (W15.X1+W25.X2)-θ5

In the process of training the neural network system 100, a large amount of known input data and output data are first used as training input data and output data, and the neural network system 100 is trained to mold each weight. And each bias value. Traditional gradient descent algorithms have traditionally been employed to correct individual weights. After molding each weight and each bias value, the neural network system 100 can be utilized to predict the output data with instant input data.

The second figure is an architectural diagram of a neural network-based reservoir water level prediction system according to the present invention. As shown, the reservoir water level prediction system 240 includes a prediction module 250, a user operation interface 260, and a database 270. The prediction module 250 includes an input data storage module 256 for capturing and storing at least one input data by a plurality of servers, and the input data includes at least real-time data and historical data of the reservoir water level; The module 252 molds at least one weight and at least one bias value according to the historical data stored in the input data storage module 256; and a prediction data operation module 254, according to the neural network learning module The weight and the bias value of the group 252 are simulated, and the real-time data is calculated to obtain at least one forecast data, and the forecast data includes at least the forecast data of the reservoir water level; the user operation interface 260 includes An operation module 262 is configured to receive a user operating condition, where the user operating condition includes at least a predicted data and/or historical data selected by the user; and a prediction The data display module 264 is configured to: retrieve, according to the user operating conditions, the predicted data and/or historical data of the selected browsing to the database 270, and display the predicted and/or historical data of the selected browsing for providing User browsing.

The prediction module 250 trains the neural network learning module 252 by a large amount of training data 200, thereby molding each weight and each bias value. Next, by input The data storage module 256 periodically collects and stores various types of real-time data from one or more data servers 220, and then uses the data stored in the input data storage module 256 as input data by the predictive data computing module 254. And according to the weights and the partial weights of the neural network learning module 252, the prediction data is further calculated, and the prediction data is stored through the database 270. The input data used by the prediction data calculation module 254 in performing the prediction calculation may further include data input by the user through the operation module 262 in addition to the data stored in the input data storage module. The user can determine the predicted data and/or historical data range to be displayed through the operation module 262. Then, the user operation interface 260 will retrieve the predicted data from the database 270 according to the predicted data range selected by the user. / or historical data, and through the forecast data display module 264 to present the predicted data and / or historical data to the user.

Please refer to the third figure and refer to the fourth figure. The third figure is the flow chart of the reservoir water level prediction method based on the neural network. The fourth picture is the water level prediction of the reservoir based on the neural network. A neural network model diagram of a particular embodiment of the system. The invention is based on a neural network-based reservoir water level prediction method, which comprises the following steps: First, in step 310, a user first constructs a neural network model, which includes determining input data of each input node, wherein the neural network The road model can be a single layer neural network model or a multilayer neural network model.

In this embodiment, at step 310, the constructed neural network model is a neural network model 400 such as shown in the fourth figure, and the neural network model 400 is used to perform, for example, a 48-hour reservoir water level forecast. As shown, the neural network model 400 is a three-layer neural network model in which the number of hidden layer neurons in the first layer of neural network 408a is For 5 to 20, the number of hidden layer neurons of the second layer neural network 408b may be 5 to 20, and the number of hidden layer neurons of the third layer neural network 408c may be 5 to 20.

It should be understood that the number of hidden layer neurons of each layer of neural network is merely exemplified herein, and the present invention is not limited to the use of 5-20 hidden layer neurons, and any number of hidden layer neurons may be set as needed.

In addition, in the embodiment, at step 310, the power generation amount data 402, the reservoir water level data 404, and the flood discharge amount data 406 are used as input data of the first layer type neural network 408a, and the reservoir inflow data 412 is used as the first Output data of a layer of neural network 408a. The input data of the second layer neural network 408b is the reservoir inflow data 412, the catchment rainfall station observation data 424, and the catchment weather forecast data 426, and the output data of the second layer neural network 408b is the reservoir inlet. Traffic prediction data 432. The input data of the third layer neural network 408c is the reservoir inflow prediction data 432, the power generation data 402 and the flood discharge data 406, and the output data of the third layer neural network 408c is the reservoir water level prediction data 442.

After step 310 is completed, step 320 is performed. At step 320, the user further determines which historical data is used as training data for training each layer of neural network, and thereby molding each layer of neural network models. Weight and bias value.

In the embodiment, the historical data is used as the training input data of the first layer type neural network 408a from the power generation data 402 of the A reservoir from 2012 to 2013, the reservoir water level data 404, and the flood discharge amount data 406, and From 2012 to 2013, the reservoir inflow data 412 of the A reservoir was used as training output data for the first layer of neural network 408a. From 2012 to 2013, the reservoir inflow data 412 of the Jia reservoir, the observation data of the rainfall station of the catchment area 424 and the meteorological forecast of the catchment area Material 426 is used as training input data for the second layer type neural network 408b, and the reservoir inflow prediction data 432 of the A reservoir from 2012 to 2013 is used as the training output data of the second layer type neural network 408b. From 2012 to 2013, the reservoir inflow forecast data 432, power generation data 402 and flood discharge data 406 of the A reservoir are used as training input materials for the third layer neural network 408c, and then from 2012 to 2013. The reservoir water level prediction data 442 of the reservoir is used as training output data for the third layer neural network 408c. Wherein, the data time point of the reservoir inflow prediction data 432 for the training of the second layer neural network 408b is compared with the observation data 424 of the catchment rainfall station for training of the second layer neural network 408b, and training The data time of the catchment weather forecast data 426 is delayed by 1 to 48 hours to simulate the relationship between the input data and the output data of the second layer neural network 408b, which is the relationship between the input data and the predicted data. The data time point of the reservoir water level prediction data 442 for the training of the third layer neural network 408c is higher than that of the training data for the training of the third layer neural network 408c, and the data time of the flood discharge data for training. The relationship between the input data and the output data of the third-layer neural network 408c is simulated by 1 to 48 hours, which is the relationship between the input data and the predicted data.

After determining the historical data as the training material 200 for training the neural networks of the various layers, proceeding to step 330, at step 330, performing the hierarchical neural networks of the neural network learning module 252 with the determined historical data. Training to mold the various weights and individual bias values of each layer of neural networks.

In this embodiment, at step 330, the neural network model 400 is trained by using the training input data and the training output data of the above 2012 to 2013 to mold the first layer neural network. Road 408a, second layer neural network 408b and third layer neural network The weights and partial weights of 408c.

Then, step 340 is performed. At step 340, the input data storage module 256 periodically retrieves the required real-time data from the data server 220 and stores the data. The captured data can also be used as a training history in the future. Information and/or as historical information for the user to view. The captured real-time data can be simultaneously stored in the database 270 for use by the user operation interface 260.

In this embodiment, at step 340, the water level data, the power generation data, and the rainfall station data of the reservoir catchment area are retrieved from the input data storage module 256 to the A server every hour at an interval of one hour. And store, go to the B server to collect the meteorological forecast data of the catchment area and store it. And when the predictive data storage module 254 performs the predictive calculation, the data is provided for the reservoir to predict the water level of the reservoir for 1 to 48 hours.

It should be understood that the above embodiments are merely examples, and the present invention is not limited to performing data capture every hour, but for several minutes to several days as the execution interval of data retrieval. At the same time, the present invention is not only used to predict reservoir water levels after 1 to 48 hours, but reservoir predictions of at least hours to days, weeks, months, and years may be performed as needed.

Next, step 350 is performed. At step 350, the prediction data calculation module 254 performs prediction operations by using the weights and the partial weight values molded by the neural network learning module 252, and the system is stored in the input data. The data of the storage module 256 is used as input data, and the predicted data is calculated for the input data, and then the calculated predicted data is stored in the database 270.

In this embodiment, the first layer-like neural network is molded by the above molding. The weights and partial weights of the 408a, the second layer neural network 408b, and the third layer neural network 408c are used to predict the water level of the A reservoir by the neural network model 400, and the predicted data is stored in the database. 270. The input data for prediction is obtained from the data collected and stored by the input data storage module 256 on the A server and the B server, that is, the method for predicting the water level of the reservoir based on the neural network is completed.

The fifth figure is a user operation display flow chart of the neural network-based reservoir water level prediction system of the present invention. The user operation display flow chart 500 starts at step 510, and the operation module 262 confirms whether a user is a legitimate user. . In one embodiment, the user enters an account password for confirmation by the operating module 262.

After confirming that the user is legally logged in, the operation module 262 is provided by the operation module 262 to use the operation interface, and at step 530, the operation module 262 receives the operation condition input by the user. The operating conditions may include, but are not limited to, the range of data selected by the user and/or some of the input data (for example, flooding simulation data and/or power generation simulation data) for prediction. The information selected by the user may be predicted data and/or historical data.

Then, in step 540, based on the operating conditions entered by the user, the predicted data display module 264 to the database 270 and/or the input data storage module 256 retrieves the predicted data and/or the user selected by the user. Selected historical data for browsing. At step 550, the predicted data display module 264 displays the data retrieved from the database 270 and/or the input data storage module 256 for viewing by the user. Displaying the data from the database 270 and/or the input data storage module 256 may be used to directly display the value of the data and/or graphically display the data. Show. The graphic display may be displayed in the form of a line graph and/or a bar graph, but is not limited thereto.

The sixth figure is a schematic diagram of a specific embodiment of the prediction result display of the reservoir-based water level prediction system based on the neural network. As shown in the figure, after the user selects the predicted data range to be browsed, the prediction data display module 264 That is, according to the predicted data range selected by the user, the corresponding forecast data is retrieved from the database 270. And in the graphic display area 610, the user-introduced reservoir inflow forecasting line graph 614 and the reservoir water level prediction line graph 616 are displayed in the form of a line graph, wherein the corresponding vertical axis 612 of the reservoir inflow forecasting line graph 614 indicates the reservoir prediction. Inflow value (CMS), horizontal axis 619 indicates the predicted time point. The corresponding vertical axis 618 of the reservoir water level prediction line chart 616 indicates the predicted water level height (m) of the reservoir, and the horizontal axis 619 indicates the predicted time point. The graphical display area 620 displays the catchment rainfall prediction bar graph 626 in the form of a bar graph, with the corresponding vertical axis 622 indicating the predicted rainfall amount (mm) and the horizontal axis 624 indicating the predicted time point.

It should be understood that the sixth figure is only a schematic diagram, and the final display graphic of the present invention is not limited thereto. For example, the graphic display area 610 can also display the reservoir inflow prediction data, the reservoir inflow historical data, and the reservoir water level prediction data. And reservoir water level history data. For another example, the graphic display area 620 can also have two sets of vertical axes according to the situation, and can display the predicted data and/or historical data of the browsing selected by the user in the form of a line graph.

200‧‧‧ Training materials

220‧‧‧Data Server

240‧‧‧ Reservoir Water Level Prediction System

250‧‧‧ Prediction Module

252‧‧‧ Neural Network Learning Module

254‧‧‧ Forecast Data Calculation Module

256‧‧‧Input data storage module

260‧‧‧User interface

262‧‧‧Operating module

264‧‧‧ Forecast data display module

270‧‧‧Database

280‧‧‧Users

Claims (13)

A reservoir-based water level prediction system based on a neural network includes: a prediction module comprising: an input data storage module, capturing and storing at least one input data from a plurality of servers, wherein the input data includes at least power generation Data, flood discharge data, observation data of rainfall stations in catchment areas, real-time data and historical data of one or part of meteorological forecast data, real-time data and historical data of reservoir water level; a type of neural network learning module, according to The historical data stored in the input data storage module is molded to at least one weight and at least one bias value; and a predictive data computing module is molded according to the neural network learning module The weight and the bias value, the real-time data is calculated to obtain at least one forecast data, and the forecast data includes at least a forecast data of a reservoir water level; a database for receiving and storing the forecast data and/or the input data And a user operation interface, receiving at least one user operating condition, and displaying the predicted data and/or according to the user operating condition Enter the information and graphics of the predicted / or the forecasts. For example, the neural network-based reservoir water level prediction system described in claim 1 wherein the neural network learning module has a first layer, a second layer, and a third layer neural network. For example, the neural network-based reservoir water level prediction system described in claim 2, wherein the neural network learning module molds the first layer neural network according to at least the historical data of the reservoir water level. The weighting and the biasing value, the predictive data computing module calculates the inflow data of the reservoir based on the weight of the first layer of the mold and the bias value, and calculates the inflow data of the reservoir water level, and the inflow of the reservoir Data for the input of the second-level neural network One of the materials. The neural network-based reservoir water level prediction system according to claim 3, wherein the neural network learning module molds the second layer neural network according to at least the reservoir inflow data. The weight and the bias value, the predictive data operation module calculates the reservoir inflow based on the weight of the second layer of the model and the bias value to calculate the instantaneous output data of at least the first layer of neural network The data, and the forecast data of the reservoir inflow is one of the input data of the third-level neural network. For example, the neural network-based reservoir water level prediction system described in claim 4, wherein the neural network learning module molds the third layer neural network according to at least the reservoir inflow prediction data. The weighting and the bias value, the predictive data computing module calculates the water level of the reservoir based on the weight of the third layer molding and the bias value for at least the instantaneous output data of the second layer neural network. Forecast data. For example, the neural network-based reservoir water level prediction system described in claim 3, wherein the neural network learning module molds the first layer neural network according to historical data of power generation and flood discharge. The weighting and the bias value, the predictive data computing module calculates the current data of at least the water level, the power generation amount and the flood discharge amount of the reservoir based on the weight of the first layer molding and the bias value. Flow data. For example, the neural network-based reservoir water level prediction system described in claim 4, wherein the neural network learning module molds the historical data according to the rainfall station observation of the catchment area and the meteorological forecast of the catchment area. The weight of the second-level neural network and the bias value, the predictive data operation module is based on the weight of the second layer of the mold and the bias value for the immediate output of at least the first-level neural network The data, the rainfall station observation in the catchment area and the real-time data of the meteorological forecast in the catchment area are used to calculate the forecast data of the reservoir inflow. For example, the neural network-based reservoir water level prediction system described in claim 5, wherein the neural network learning module molds the third layer neural network according to historical data of power generation and flood discharge. The weighting and the bias value, the predictive data computing module is based on the third The weight of the layer molding and the bias value calculate the predicted data of the reservoir water level by calculating the instantaneous data of the instantaneous output data, the power generation amount and the flood discharge amount of the second layer neural network. For example, the neural network-based reservoir water level prediction system described in claim 1 includes the reservoir inflow prediction data. A reservoir-based water level prediction method based on a neural network includes: constructing a neural network model; determining historical data for training; modeling the weights and bias values of the neural network based on the training historical data; collecting at least An instant input data, the real-time input data includes at least the power generation data, the flood discharge data, the observation data of the catchment rainfall station, the real-time data of one or part of the meteorological forecast data, and the real-time data of the reservoir water level; The weight and the bias value of the neural network are simulated, and the real-time input data is calculated to obtain at least one prediction data, and the prediction data includes at least the reservoir water level prediction data. For example, the neural network-based reservoir water level prediction method described in claim 10, wherein the neural network model has a plurality of layers of neural networks. For example, the neural network-based reservoir water level prediction method described in claim 11 wherein at least one layer of neural network calculation output data is reservoir inflow prediction data, and the reservoir inflow prediction data is the layer. One of the input data of a class of neural networks below the neural network. For example, the neural network-based reservoir water level prediction method described in claim 10, wherein the prediction data includes reservoir inflow prediction data.