A proxy decision-making method for water supply system scheduling

Abstract

A method for scheduling proxy decision-making of a water supply system, comprising the following steps: 1) collecting measured monitoring data of a water delivery project, including water level, flow rate, and gate opening and closing conditions. Dividing sample data into a training set, a test set, and a prediction set. 2) Constructing a BP neural network model, which is composed of an input layer, a hidden layer, and an output layer. 3) Inputting sample data of the training set into the BP neural network model, the training process includes an information propagation process and an error propagation process. 4) Outputting a control decision. In step 3), a relatively ideal BP neural network model has been obtained, and the output result of the output layer is in the interval [0,1]. The output result is denormalized to obtain the gate opening value, and the gate control decision is output. The present invention can reflect the mutual connection and mutual influence between the various elements of the water delivery system; the valve opening value can be predicted by elements such as water level and target flow rate, and the result is reliable, which can provide a reference for actual scheduling management; making water transfer decisions more intelligent and accurate, and facilitating the operation and management of actual projects.

Description

Technical Field

[0001] The invention belongs to the field of long-distance water transmission scheduling and relates to a water supply system scheduling agent decision method. Background Art

[0002] In order to solve the problem of uneven spatial distribution of water resources, my country has built a large number of long-distance water transfer projects in recent years. The operation and dispatching mode of water transfer projects, as a key link affecting the operation efficiency and structural safety of the project, directly determines the project benefits. Therefore, in order to fully realize the construction goals of water transfer projects and ensure the safety of project operation, it is necessary to study the operation and dispatching of long-distance water transfer projects. At present, open channel control algorithms are mostly used in China to replace manual experience for precise open channel regulation. The optimization scheduling model is mainly constructed based on the simulation results of the open channel simulation model to optimize the control strategy. However, the calculation of the optimization model takes a long time, especially when facing larger-scale and more complex topological water transfer projects. The calculation time will be longer. In addition, most optimization models are based on the Saint-Venant model or its derivative models. Affected by the complexity of water flow operation, changes in physical parameters of the water transfer system, measurement errors and other factors, the accuracy of the constructed model will change with the operation time of the project, so this type of method is not suitable for application in actual projects. Therefore, while improving the efficiency of water transfer projects, giving full play to the potential of projects, improving project benefits and ensuring project safety, how to choose the optimal scheduling control scheme and solve the long-distance water transfer system regulation strategy has become an urgent problem to be solved. Summary of the invention

[0003] In order to respond more quickly to the centralized control purpose of the water transmission system and meet the user's demand for "timely and appropriate" water transmission, the present invention provides a water supply system scheduling agent decision method to replace the optimization scheduling model to predict the valve scheduling plan of the water transmission system. The method uses the BP neural network to construct a long-distance water transmission system operation scheduling model, which reflects the mapping relationship between the water level, flow and valve opening of the water transmission system, and determines the water transmission system control strategy under different target conditions.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is:

[0005] A water supply system scheduling agent decision-making method comprises the following steps:

[0006] Step 1: Sample data acquisition and processing

[0007] Collect measured monitoring data of water transfer projects, including water level, flow, and gate opening and closing. Assuming that the valve opening remains unchanged during the period, the sample data is processed as the daily average water transfer flow. According to the requirements of building a BP neural network model, the sample data is divided into a training set data group, a test set data group, and a prediction set data group, and each set of data is obtained by randomly combining the sample data. Due to the different ranges and dimensions of different data, all sample data need to be normalized before training the neural network model. The normalization formula is formula (1-1), and the normalization interval is [0, 1].

[0008]

[0009] Where: x is the original value of the sample data; x' is the normalized value corresponding to x; x min is the minimum value in the sample data; x max is the maximum value in the sample data.

[0010] Step 2: Construct BP neural network model

[0011] The present invention adopts a three-layer BP neural network, which is composed of an input layer, a hidden layer, and an output layer. Its characteristic is that various neurons between the upper and lower layers are fully connected, that is, each unit in the lower layer is weightedly connected with each unit in the upper layer, and there is no connection between the units in each layer. The transfer function adopts an S-type (sigmoid) function. Since the valve opening is related to its initial state, water level and target water demand (water delivery), the water delivery flow of each user in the current state, the water level value at the entrance of the water delivery project within the time period, the head value after each valve within the time period, and the daily water demand flow of each user can be used as input factors of the input layer. The hidden layer is a nonlinear relationship between the gate opening, flow, and water level. Since the number of hidden layer nodes determines the network structure, it affects the accuracy and computational complexity of the neural network recognition. If the number of hidden layer nodes increases, although the network recognition accuracy is improved, it will also cause the network structure to be more complex and increase the computational time. Therefore, the selection of the number of hidden layer nodes of the neural network is very important. The hidden layer node selection method adopted by the present invention is the empirical formula method (1-2). The output content of the output layer is the gate opening.

[0012]

[0013] Where: m is the number of nodes in the hidden layer; n is the number of nodes in the input layer; p is the number of nodes in the output layer; α is a constant between 1 and 10.

[0014] Step 3: Train the BP neural network model

[0015] The normalized training set sample data is input into the BP neural network model to start training. The BP neural network training process includes information propagation process and error propagation process. The information propagation process is the transmission of input layer information to the output layer through the hidden layer, and the error propagation process is the error between the expected output of the output layer and the network output value is transmitted back to the input layer through the hidden layer.

[0016] In the process of information transmission, there are errors in each layer of neurons. In order to ensure that the error between the network output value and the expected output value is reduced, the connection weight value between each layer of neurons needs to be continuously adjusted until the error is less than the target value. The training set data trained by the BP neural network is fitted with the original sample data. The present invention uses the mean square error between the BP neural network prediction result and the actual result of the output layer to measure and evaluate the quality of the scheduling function calculation result. The sample mean square error MSE (Mean-Square Error) needs to be controlled within 0.02. During the fitting process, if the MSE is greater than 0.02, the result is considered unreasonable, and a new round of training is started to adjust the weight value between each neuron until the result is reasonable; if the MSE is less than 0.02, it is considered that the fitting result is good and the modeling result is ideal, indicating that the BP neural network model can more accurately obtain the nonlinear relationship between the valve opening value and the water level and target flow.

[0017] The normalized test set data is input into the trained BP neural network model for testing to verify the accuracy of the model. The test set data predicted by the BP neural network is fitted with the actual test set result data. If the MSE is less than 0.02, the model is considered to be ideal; if the MSE is greater than 0.02, the model is considered to be unsatisfactory and needs to be retrained with more training set data.

[0018] Step 4: Output control decision In step 3, a relatively ideal BP neural network model has been obtained. The output result of the output layer is in the interval [0,1]. The output result is denormalized and restored to the original dimension to obtain the gate opening value, and the gate control decision can be output. The denormalization formula is (1-3).

[0019] x=x'×(x max -x min )+x min (1-3)

[0020] Where: x is the original value of the sample data; x' is the normalized value corresponding to x; x min is the minimum value in the sample data; x max is the maximum value in the sample data.

[0021] The beneficial effects of the present invention are:

[0022] The neural network model established by the present invention can refine the operating rules of the water delivery system and reflect the mutual connection and mutual influence between the various elements of the water delivery system. The model can be used to predict the valve opening value through factors such as water level and target flow. The result is relatively reliable and can provide a reference for actual scheduling management. The neural network model requires less time for simulation calculation, and can respond to emergencies quickly and provide solutions. As the operating time increases, the actual monitoring data gradually increases. The monitoring data can be used as sample data to construct a neural network model to refine the actual water delivery operation scheduling rules, make water diversion decisions more intelligent and accurate, and facilitate the operation and management of the actual project. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a flow chart of the present invention.

[0024] Figure 2 This is the test result of valve #1.

[0025] Figure 3 This is the test result of valve 2#. DETAILED DESCRIPTION

[0026] The present invention is further described below with reference to a specific embodiment of the Guanyinge Reservoir in Liaoning Province.

[0027] The Guanyin Pavilion Reservoir Water Conveyance Project is mainly arranged between Guanyin Pavilion Reservoir and Benxi City. It consists of 1 main pipeline, 2 branch pipelines and 6 branch pipelines. The entire line involves 1 power station, 1 surge tank, 1 labyrinth weir and 2 water distribution stations (including 10 flow regulating valves in total). The total length of the main line is 48km, and the maximum water diversion flow is 125*104m 3 / d. The project adopts a water delivery method that combines pressurized flow and non-pressurized flow. The main pipeline and branch pipeline mainly use non-pressurized tunnel water delivery, while the branch line mainly uses pressurized gravity flow water delivery.

[0028] At present, the water supply scope of Guanyin Reservoir Water Conveyance Project is 1 main pipeline, 2 branch pipelines and 3 branch pipelines. The actual water diversion flow of the main pipeline is less than the design flow. The water supply project mainly adjusts the water volume of the entire line by controlling the 3 flow regulating valves in the water distribution station to meet the water supply goals of 3 users (water company, Xincheng, Qiaobei Water Plant). The project currently has 5 flow meters installed at the entrance of the main pipeline (power station outlet) and about 8m from the entrance of the 3 branch pipelines (before the flow regulating valve in the water distribution station); the project currently has 4 pressure gauges installed at the entrance of the main pipeline (power station outlet) and about 8m from the entrance of the 3 branch pipelines (before the flow regulating valve in the water distribution station). The data collection interval is 1 hour.

[0029] 1) Data acquisition and processing.

[0030] Due to the lack of measured data, it is necessary to first build a hydraulic model for simulation, accurately simulate the water flow process inside the water delivery system, obtain 400 sets of simulated sample data, and carry out follow-up work. Then, according to the needs of building a neural network model, the sample data is divided into three groups, 288 groups as training set data, 72 groups as test set data, and 40 groups as prediction set data, and each data set is obtained by random combination from the sample data. The sample data is normalized and the output results are denormalized.

[0031] 2) Construct a BP neural network model.

[0032] The Guanyin Pavilion Reservoir Water Conveyance Project in Liaoning Province currently has three operating gates. However, in the actual process, since the line where the 3# gate is located contains a flow weir, its flow rate is affected by the water head in front of the weir, and the water head in front of the weir is mainly controlled by the 1# gate and the 2# gate. Therefore, the change in the opening of the 3# gate has almost no effect on the water diversion value within a certain range, and there is a non-corresponding relationship between its opening value and the water diversion flow. Therefore, the output factor does not consider the 3# gate, but only the 1# gate and the 2# gate. The neural network model is constructed for the 1# gate and the 2# gate respectively for prediction. The example adopts a 3-layer neural network structure, with 10 nodes in the input layer, 12 nodes in the hidden layer, and 1 node in the output layer.

[0033] 3) Train the BP neural network model.

[0034] The normalized training set data is input into the BP neural network and trained using the parameters set in step 2. The training set result data trained by the neural network is fitted with the original sample result data. The mean square error of the opening of gate 1# and gate 2# is less than 0.02, which meets the accuracy requirements and the result fits well. The normalized test set data is input into the trained model, and the test set results are fitted with the original sample results. The mean square error of the opening of gate 1# and gate 2# is less than 0.02, which meets the accuracy requirements. It is considered that the model is ideal.

[0035] Table 1 shows the training results of BP neural network. Figure 2 and Figure 3 They are the comparison results of the simulation results of Gate 1# and Gate 2# output by the neural network model and the corresponding data in the original sample test set, with a total of 72 test data.

[0036] Table 1 Neural network training results

[0037]

[0038] from Figure 2 , Figure 3It can be seen that although the gate openings at individual sample points vary greatly, the overall result has a good fit; the simulation results of the neural network model are consistent with the change trend of the original data sample results, indicating that the water delivery system operation and scheduling model constructed using the BP neural network can more accurately reflect the gate operation control law, the modeling results are relatively ideal, and the model output results are of practical significance for guiding the operation and scheduling management of the water delivery system.

[0039] 4) Output regulatory decisions.

[0040] The result in step 3 has a good fitting degree, and a relatively ideal BP neural network model is obtained. After the output result is denormalized, the gate opening value is obtained, from which a gate control plan can be formulated to schedule the operation of the water delivery system. The above-mentioned embodiments only express the implementation mode of the present invention, but they cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that for those skilled in the art, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A water supply system scheduling agent decision method, characterized in that: The following steps are involved: Step 1: Sample data acquisition and processing Collect measured monitoring data of water transfer projects, including water level, flow, and gate opening and closing conditions; assume that the valve opening remains unchanged during the period, and process the sample data as the daily average water transfer flow; according to the requirements of building a BP neural network model, divide the sample data into a training set data group and a test set data group, and each set of data is obtained by randomly combining the sample data; Step 2: Construct BP neural network model A three-layer BP neural network is adopted, which consists of an input layer, a hidden layer, and an output layer. Its characteristic is that various neurons between the upper and lower layers are fully connected, that is, each unit in the lower layer is weightedly connected with each unit in the upper layer, and there is no connection between the units in each layer; the transfer function adopts an S-type function; since the valve opening is related to its initial state, water level and target water demand, the water flow of each user in the current state, the water level value at the entrance of the water supply project within the time period, the head value after each valve within the time period, and the daily water demand flow of each user are used as input factors of the input layer; the hidden layer is the nonlinear relationship between the gate opening, flow and water level; the hidden layer node selection method adopted is the empirical formula method (1-2); the output content of the output layer is the gate opening; Where: m is the number of nodes in the hidden layer; n is the number of nodes in the input layer; p is the number of nodes in the output layer; α is a constant between 1 and 10; Step 3: Train the BP neural network model Input the training set sample data after normalization in step 1 into the BP neural network model to start training. The BP neural network training process includes information propagation process and error propagation process; The information propagation process is the transmission of the input layer information to the output layer through the hidden layer, and the error propagation process is the transmission of the error between the expected output of the output layer and the network output value through the hidden layer to the input layer in reverse. In the process of information transmission, there are errors in each layer of neurons. In order to ensure that the error between the network output value and the expected output value is reduced, the connection weight value between the neurons in each layer needs to be continuously adjusted until the error is less than the target value; the training set data trained by the BP neural network is fitted with the original sample data, and the mean square error between the BP neural network prediction result and the actual result of the output layer is used to measure and evaluate the quality of the scheduling function calculation result. The sample mean square error MSE needs to be controlled within 0.02; in the fitting process, if MSE is greater than 0.02, the result is considered unreasonable, and a new round of training is started to adjust the weight value between each neuron until the result is reasonable; if MSE is less than 0.02, it is considered that the fitting result is good and the modeling result is ideal, indicating that the BP neural network model can more accurately obtain the nonlinear relationship between the valve opening value and the water level and target flow; Input the normalized test set data into the trained BP neural network model for testing to verify the accuracy of the model; The test set data predicted by the BP neural network is fitted with the actual test set result data. If the MSE is less than 0.02, the model is considered to be ideal; if the MSE is greater than 0.02, the model is considered to be unsatisfactory and needs to be retrained with more training set data. Step 4: Output control decision. In step 3, a relatively ideal BP neural network model has been obtained. The output result of the output layer is in the interval [0,1]. The output result is denormalized and restored to the original dimension to obtain the gate opening value, and the gate control decision can be output.

2. A water supply system scheduling agent decision method according to claim 1, characterized in that: In the step 1, due to the different ranges and dimensions of different data, all sample data need to be normalized before training the neural network model. The normalization formula is formula (1-1), and the normalization interval is [0, 1]; Where: x is the original value of the sample data; x' is the normalized value corresponding to x; x min is the minimum value in the sample data; x max is the maximum value in the sample data.

3. A water supply system scheduling agent decision method according to claim 1, characterized in that: In the step 4, the denormalization formula is (1-3); x=x'×(x max -x min )+x min (1-3) Where: x is the original value of the sample data; x' is the normalized value corresponding to x; x min is the minimum value in the sample data; x max is the maximum value in the sample data.

Images