An energy-saving control method for parallel water pumps in a SHMFF pumping system

Abstract

The present invention relates to an energy-saving control method for parallel water pumps in an SHMFF pumping system. Compared with the prior art, this method solves the problem of difficulty in achieving coordinated energy-saving control of pump groups. The present invention includes the following steps: obtaining the operating parameters of the parallel water pump group; constructing a BP neural network; training the BP neural network; optimizing the BP neural network A and the BP neural network B; generating an optimal flow distribution scheme; and energy-saving control of the parallel water pumps. The present invention utilizes a limited data set tested in actual use to achieve energy-saving control optimization for parallel variable-frequency water pump groups with different performances. It can be tailored to the actual operating conditions of the equipment and calculated to maximize the energy-saving effect of the variable-frequency pump group.

Description

Technical Field

[0001] The present invention relates to the technical field of pumping system control, and in particular to an energy-saving control method for parallel water pumps in an SHMFF pumping system. Background Art

[0002] The Steady-State High Magnetic Field Facility (SHMFF), a major scientific and technological infrastructure facility constructed during my country's 11th Five-Year Plan, successfully completed acceptance on September 27, 2017. Its magnet technology and overall performance are internationally leading. The pumping system is crucial to the operation of this large-scale national experimental facility. The SHMFF's five water-cooled magnets generate tens of megawatts of heat during magnet excitation, requiring deionized water for cooling.

[0003] To meet the water supply needs of different magnets, the pumping system has a wide flow rate range, requiring multiple water pumps to operate in parallel. These pumps include two types: the MC100-300 centrifugal water pump and the Multitec A-125 centrifugal water pump. Due to the unique structure of the magnets, it is difficult to accurately calculate their local resistance losses when designing the magnet water cooling system. The parallel pump group has too small a design margin, causing the water pumps to operate beyond their best efficiency point (BEP), resulting in rapidly increasing noise and vibration, and causing motor overload. This not only severely consumes energy but also increases the pump failure rate, affecting the normal progress of magnet experiments. In severe cases, the magnet coils can be damaged due to insufficient cooling.

[0004] Existing energy-saving control methods for parallel pump groups typically use factory sample data or operating data under rated operating conditions measured on a laboratory test platform to fit performance curves and perform modeling calculations. This approach is suitable for engineering design or factory experiments. However, after long-term use, pump performance degrades to varying degrees compared to factory data due to mechanical wear, electrochemical corrosion, and varying usage frequencies. This results in significant deviations between actual operating data and factory data. Furthermore, in actual use, due to limitations in the actual pipe network layout, it is impossible to perform performance testing of the pumps under rated operating conditions, making it impossible to establish an accurate mathematical model.

[0005] Therefore, based on the current situation where the hydraulic and structural design of the existing pump group is difficult to change, how to improve the reliability and energy-saving effect of the pump group operation through the optimization of operation scheduling and control strategies has become a technical problem that needs to be solved urgently. Summary of the Invention

[0006] The purpose of the present invention is to solve the defect in the prior art that it is difficult to achieve coordinated energy-saving control of pump groups, and to provide an energy-saving control method for parallel water pumps in a SHMFF pumping system to solve the above problem.

[0007] In order to achieve the above object, the technical solution of the present invention is as follows:

[0008] A method for energy-saving control of parallel water pumps in a SHMFF pumping system, comprising the following steps:

[0009] Acquisition of operating parameters of the parallel water pump group: Acquisition of operating data of the MC100-300 centrifugal water pump and the Multitec A-125 centrifugal water pump in the SHMFF pumping system and pre-processing;

[0010] Construction of BP neural network: BP neural network A is constructed for MC100-300 centrifugal water pump, and BP neural network B is constructed for MultitecA-125 centrifugal water pump;

[0011] BP neural network training: BP neural network A and BP neural network B are trained using the measured data of MC100-300 centrifugal water pump and Multitec A-125 centrifugal water pump respectively;

[0012] Optimization of BP neural network A and BP neural network B: Use genetic algorithm to optimize the weights and thresholds of BP neural network A and BP neural network B respectively, and obtain the optimized neural network GA-BP A and neural network GA-BP B;

[0013] Generation of the optimal flow distribution plan: Using the particle swarm optimization method to connect the neural network GA-BP A and the neural network GA-BP B, the optimal flow distribution plan is searched for each pump combination under the given demand.

[0014] Energy-saving control of parallel water pumps: The optimal flow distribution plan, that is, the single pump head and the searched optimal flow distribution are input into the neural network GA-BP A or neural network GA-BP B corresponding to the corresponding water pump model. The total power of different pump groups after parallel connection is calculated, and the one with the lowest total power is selected as the optimal parallel operation plan, thereby performing energy-saving control on the parallel water pumps.

[0015] The acquisition of the operating parameters of the parallel water pump group includes the following steps:

[0016] The head, flow rate, and shaft power values ​​at the operating point of the MC100-300 centrifugal water pump group were measured using the pressure gauges, flow meters, and torque and speed sensors on the pump shafts when the pumps were operated in variable frequency mode, respectively. These values ​​are recorded as Dataset A.

[0017] The head, flow rate, and shaft power values ​​at the operating point of the Multitec A-125 centrifugal water pump group were measured using the pressure gauges, flow meters, and torque and speed sensors on the pump shafts when the pumps were operated in variable frequency mode. These values ​​are recorded as Dataset B.

[0018] The head, flow rate and shaft power data measured above are normalized and preprocessed. The data normalization formula is as follows:

[0019] x k =(x k -x min ) / (x max -x min ),

[0020] Among them, x k is the head, flow or shaft power value, x min is the minimum value of the corresponding data set, x max is the maximum value of the corresponding data set.

[0021] The construction of the BP neural network comprises the following steps:

[0022] Establish BP neural network A for MC100-300 centrifugal water pump;

[0023] The BP neural network A is assumed to include an input layer, a hidden layer, and an output layer. The input layer has two nodes, representing the head and flow rate of a single pump, and the output layer has one node, representing the shaft power of a single pump. The number of hidden layers is calculated using the formula 2n+1, where n is the number of input nodes.

[0024] The hidden layer is set to use the tangent S-type transfer function, and the input of the hidden layer neurons is the weighted sum of all inputs, that is,

[0025] x j =∑ i w ij X i ,

[0026] Output x′ of hidden layer neurons j Use S function to excite x j ,have to

[0027]

[0028] but

[0029]

[0030] Among them, x j is the input of the hidden layer neurons, w ij is the connection weight between the input layer and the hidden layer, x′ j is the output of the hidden layer neurons;

[0031] Set the output layer to use a linear transfer function, and the output layer neuron Y k The output is:

[0032]

[0033] Among them, the error between the network k-th output and the corresponding ideal output is e k , the error performance index function of the pth sample is N is the number of network output layers;

[0034] Establish BP neural network B for MultitecA-125 centrifugal water pump.

[0035] The training of the BP neural network comprises the following steps:

[0036] Using the obtained data set A, 2 / 3 of the data is used to train the neural network A, and the other 1 / 3 of the data is used as the validation set of the BP neural network A;

[0037] Use the obtained data set B, 2 / 3 of the data is used to train neural network B, and the other 1 / 3 of the data is used as the validation set of BP neural network B;

[0038] The training function is set to use the Levenberg-Marquardt algorithm, and the network learning function is set to use the BP learning rule with momentum term.

[0039] The optimization of the BP neural network A and the BP neural network B comprises the following steps:

[0040] The genetic algorithm is used to reduce the fitness function value through multiple iterations of selection, crossover, and mutation operations, thereby continuously optimizing the weights and thresholds of the BP neural network A until the allowable mean square error is met:

[0041] Population initialization: The individual coding method is real number coding, and the individual coding length is 18+7=25;

[0042] Fitness function calculation: After training the BP neural network based on each set of water pump training data, the shaft power output of the system is predicted. The absolute value of the error between the predicted output and the expected output is used as the individual fitness Q. The expected output is the actual measured shaft power value. The calculation formula is as follows:

[0043]

[0044] Where N is the number of network output nodes, Y k is the expected output of the kth node of the BP neural network, O k is the predicted output of the kth node, δ is the coefficient, and abs is the absolute value;

[0045] Selection operation: Roulette method is used, and the selection strategy is based on the fitness ratio. The selection probability p of each individual i is i for

[0046]

[0047] Where Q i and Q j is the fitness value of the individual, δ is the coefficient, and S is the number of individuals in the population;

[0048] Crossover operation: Using real number crossover method, the mth chromosome a m and the nth chromosome a n The crossover operation method at position j is as follows:

[0049] a mj =a mj (1-b)+a nj b

[0050] a nj =a nj (1-b)+a mj b

[0051] Where a mj Refers to chromosome a m The jth position, a nj Refers to chromosome a n The jth bit of , b is a random number between [0,1];

[0052] Mutation operation: select the jth gene a of the i-th individual ij Perform mutation, the mutation algorithm is as follows:

[0053]

[0054] Where a max Gene a ij The upper bound of a min Gene a ij The next term, H(g)=r2(1-g / G max ) 2 , r2 is a random number, g is the current number of iterations, G max is the maximum number of evolutions, r is a random number between [0,1];

[0055] Update weights and thresholds: After multiple iterations of the genetic algorithm, the optimal individual with the lowest fitness function is obtained, and the weights and thresholds of the BP neural network are assigned to it. After optimization, the neural network GA-BP A is obtained;

[0056] According to the above methods and steps, the weights and thresholds of the BP neural network B are optimized using a genetic algorithm, and the neural network GA-BP B is obtained after optimization.

[0057] The generation of the optimal flow distribution solution includes the following steps:

[0058] Under a given water supply demand, the particle swarm optimization method is used to connect the neural network GA-BP A and the neural network GA-BP A, the neural network GA-BP A and the neural network GA-BP B, and the neural network GA-BP B and the neural network GA-BP B;

[0059] Using the single pump head and flow rate (H, Q) vector group as the particle, the particle fitness function F is as follows:

[0060]

[0061] Where L is the number of water pumps, L≤m+n, m is the total number of water pumps A, n is the total number of water pumps B, N shaft is the shaft power value of a single pump;

[0062] The number of particles is set to the number of running pumps. The lift between different particles is the same, and the sum of the flow rates is the required flow rate.

[0063] The water pump combinations are set as several MC100-300 centrifugal water pumps connected in parallel, several MultitecA-125 centrifugal water pumps connected in parallel, and several MC100-300 centrifugal water pumps and several MultitecA-125 centrifugal water pumps connected in parallel. The number of parallel operation units is less than the total number of SHNFF pumping systems.

[0064] By updating the particle velocity and position, the optimal flow distribution is searched for each pump combination so that the fitness function of the combination is the lowest and the optimal flow distribution plan is generated.

[0065] Beneficial effects

[0066] Compared with the existing technology, the energy-saving control method of parallel water pumps in an SHMFF pumping system of the present invention utilizes a limited data set tested in actual use to achieve energy-saving control optimization of parallel variable-frequency water pump groups with different performances. It can be tailored to the actual operating conditions of the equipment and calculated to maximize the energy-saving effect of the variable-frequency pump group.

[0067] The method of the present invention does not require mathematical modeling. The BP neural network based on PSO-GA predicts and selects the optimal energy-saving operation plan for the pump group according to the performance data of the equipment during actual operation. The method is more in line with the actual situation of the equipment and can be updated in real time according to the operation changes of the equipment, so as to obtain more realistic and accurate prediction results, thereby formulating a practical and effective operation plan for the parallel pump group. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 is a method sequence diagram of the present invention;

[0069] Figure 2This is a topological diagram of the BP neural network involved in the present invention;

[0070] Figure 3 This is a topological diagram of the optimized neural network GA-BP involved in the present invention;

[0071] Figure 4 This is a flow chart of the function optimization algorithm of the PSO algorithm involved in the present invention;

[0072] Figure 5 This is a graph showing changes in the mean square error during BP neural network training involved in the present invention;

[0073] Figure 6 This is a diagram of fitness changes during the process of optimizing weight thresholds using a genetic algorithm in the GA-BP neural network involved in the present invention. DETAILED DESCRIPTION

[0074] In order to provide a further understanding and appreciation of the structural features and effects achieved by the present invention, a detailed description is provided with reference to preferred embodiments and accompanying drawings as follows:

[0075] like Figure 1 As shown, the energy-saving control method of the parallel water pumps of the SHMFF pumping system described in the present invention includes the following steps:

[0076] The first step is to obtain the operating parameters of the parallel water pump group: obtain the operating data of the MC100-300 centrifugal water pump and the MultitecA-125 centrifugal water pump in the SHMFF pumping system and perform preprocessing.

[0077] (1) The head, flow rate and shaft power values ​​of the working point when single pump and multiple pumps are operated with variable frequency are measured by the pressure gauges, flow meters at the inlet and outlet of the MC100-300 centrifugal water pump group and the torque and speed sensor on the pump shaft, respectively, and recorded as data set A.

[0078] (2) The head, flow rate and shaft power values ​​of the working point when single pump and multiple pumps are operated with variable frequency are measured by the pressure gauges, flow meters and torque speed sensors on the pump shaft of the Multitec A-125 centrifugal water pump group, respectively, and recorded as data set B.

[0079] (3) The head, flow rate and shaft power data measured above are normalized and preprocessed. The data normalization formula is as follows:

[0080] x k =(x k -x min ) / (x max -x min ),

[0081] Among them, x kis the head, flow or shaft power value, x min is the minimum value of the corresponding data set, x max is the maximum value of the corresponding data set.

[0082] The second step is the construction of BP neural network: construct BP neural network A for MC100-300 centrifugal water pump and construct BP neural network B for MultitecA-125 centrifugal water pump.

[0083] Since there are multiple water pumps in the SHMFF pumping system, there are two models, namely MC100-300 and MultitecA-125. Since these two models of centrifugal water pumps correspond to multiple water pumps, it is necessary to establish the BP neural network for the two models of centrifugal water pumps using the same method.

[0084] In practical applications, the head, flow rate, and shaft power parameters of a single pump operating at different frequencies, in parallel with two, three, and four pumps, are measured. Care is taken to select a data set that covers as large an operating range as possible. It can be found that the HQ performance curves for different frequencies, fitted using factory sample data, show some deviations from the curves fitted with factory data, especially under some off-design operating conditions. Therefore, even if the fitted performance curves are accurate, they cannot describe the actual operating point parameters. BP neural networks can effectively address this problem.

[0085] The specific steps of constructing the BP neural network are as follows:

[0086] (1) Establish BP neural network A for MC100-300 centrifugal water pump.

[0087] A1) If Figure 2 As shown, the BP neural network A is set to include an input layer, a hidden layer and an output layer; wherein, the input layer has two nodes, which are the head and flow of a single pump respectively, and the output layer has one node, which is the shaft power value of a single pump. The number of hidden layers is calculated according to the formula 2n+1, where n is the number of input nodes.

[0088] A2) Set the hidden layer to use the tangent sigmoid transfer function, and the input of the hidden layer neurons is the weighted sum of all inputs, that is,

[0089] x j =∑ i w ij X i ,

[0090] Output x′ of hidden layer neurons j Use S function to excite x j ,have to

[0091]

[0092] but

[0093]

[0094] Among them, x j is the input of the hidden layer neurons, w ij is the connection weight between the input layer and the hidden layer, x′ j is the output of the hidden layer neurons.

[0095] A3) Set the output layer to use a linear transfer function, and the output layer neurons Y k The output is:

[0096]

[0097] Among them, the error between the network k-th output and the corresponding ideal output is e k , the error performance index function of the pth sample is N is the number of network output layers.

[0098] (2) Using the same method as above, a BP neural network B is established for the Multitec A-125 centrifugal water pump.

[0099] The third step is the training of BP neural network: the measured data of MC100-300 centrifugal water pump and MultitecA-125 centrifugal water pump are used to train BP neural network A and BP neural network B respectively.

[0100] (1) Using the obtained data set A, 2 / 3 of the data is used to train the neural network A, and the other 1 / 3 of the data is used as the verification set of the BP neural network A.

[0101] (2) Using the obtained data set B, 2 / 3 of the data is used to train the neural network B, and the other 1 / 3 of the data is used as the verification set of the BP neural network B.

[0102] (3) The training function is set to use the Levenberg-Marquardt algorithm, and the network learning function is set to use the BP learning rule with momentum term.

[0103] The fourth step is the optimization of BP neural network A and BP neural network B: using genetic algorithm to optimize the weights and thresholds of BP neural network A and BP neural network B respectively, and obtain the optimized neural network GA-BP A and neural network GA-BP B.

[0104] Because the BP neural network is an optimization method that searches for a local solution, the weights and thresholds are often modified gradually along the path of local optimization. This often causes the BP neural network to fall into the trap of a local optimal solution. Furthermore, a small change in the BP neural network can result in different global results. Training the network with different weights often results in convergence to different local minima, resulting in significant differences in the results and time of each training session. Genetic algorithms, on the other hand, have excellent global search capabilities and can quickly search the entire solution space without falling into the trap of rapidly declining local optimal solutions. Therefore, the present invention uses genetic algorithms to optimize the weights and thresholds of the BP neural network.

[0105] The specific steps are as follows:

[0106] (1) Use the genetic algorithm to reduce the fitness function value through multiple iterations of selection, crossover, and mutation operations, thereby continuously optimizing the weights and thresholds of the BP neural network A until the allowable mean square error is met:

[0107] B1) Population initialization: The individual coding method is real number coding, and the individual code length is 18+7=25;

[0108] B2) Fitness function calculation: After training the BP neural network based on each set of water pump training data, the system's shaft power output is predicted. The absolute value of the error between the predicted output and the expected output is used as the individual fitness function Q. The expected output is the actual measured shaft power value. The calculation formula is as follows:

[0109]

[0110] Where N is the number of network output nodes, Y k is the expected output of the kth node of the BP neural network, O k is the predicted output of the kth node, δ is the coefficient, and abs is the absolute value;

[0111] B3) Selection operation: Roulette wheel method is used, and the selection strategy is based on the fitness ratio. The selection probability p of each individual i is i for

[0112]

[0113] Where Q i and Q j is the fitness value of the individual, δ is the coefficient, and S is the number of individuals in the population;

[0114] B4) Crossover operation: Using real number crossover method, the mth chromosome a m and the nth chromosome a n The crossover operation method at position j is as follows:

[0115] a mj =a mj (1-b)+a nj b

[0116] a nj =a nj (1-b)+a mj b

[0117] Where a mj Refers to chromosome a m The jth position, a nj Refers to chromosome a n The jth bit of , b is a random number between [0,1];

[0118] B5) Mutation operation: Select the jth gene a of the i-th individual ij Perform mutation, the mutation algorithm is as follows:

[0119]

[0120] Where a max Gene a ij The upper bound of a min Gene a ij The next term, H(g)=r2(1-g / G max ) 2 , r2 is a random number, g is the current number of iterations, G max is the maximum number of evolutions, r is a random number between [0,1];

[0121] B6) Update weights and thresholds: After multiple iterations of the genetic algorithm, the optimal individual with the lowest fitness function is obtained, and the weights and thresholds of the BP neural network are assigned to it. After optimization, the neural network GA-BP A is obtained.

[0122] (2) According to the above methods and steps, the weights and thresholds of the BP neural network B are optimized using a genetic algorithm, and the neural network GA-BP B is obtained after optimization.

[0123] Step 5: Generate the optimal flow distribution plan: Use the particle swarm optimization method to connect the neural network GA-BP A and the neural network GA-BP B to search for the optimal flow distribution plan for each pump combination under given demand.

[0124] (1) Under a given water supply demand, the particle swarm algorithm is used to connect the neural network GA-BP A and the neural network GA-BP A (a combination of several MC100-300 centrifugal water pumps), the neural network GA-BP A and the neural network GA-BP B (a combination of several MultitecA-125 centrifugal water pumps), and the neural network GA-BP B and the neural network GA-BP B (a combination of multiple MC100-300 centrifugal water pumps and multiple MultitecA-125 centrifugal water pumps).

[0125] Using the single pump head and flow rate (H, Q) vector group as the particle, the particle fitness function F is as follows:

[0126]

[0127] Where L is the number of water pumps, L≤m+n, m is the total number of water pumps A, n is the total number of water pumps B, N shaft is the shaft power value of a single pump.

[0128] (2) The number of particles is set to the number of operating water pumps. The lift between different particles is the same, and the total flow rate is the required flow rate.

[0129] (3) The water pump combinations are set to be several MC100-300 centrifugal water pumps connected in parallel, several MultitecA-125 centrifugal water pumps connected in parallel, and several MC100-300 centrifugal water pumps and several MultitecA-125 centrifugal water pumps connected in parallel. The number of parallel operation units is lower than the total number of SHNFF pumping systems.

[0130] (4) Using the updated particle velocity and position, search for the optimal flow distribution for each pump combination so that the fitness function under the combination is the lowest and the optimal flow distribution plan is generated.

[0131] The sixth step is energy-saving control of parallel water pumps: the optimal flow distribution plan, that is, the single pump head and the searched optimal flow distribution are input into the neural network GA-BP A or neural network GA-BP B corresponding to the corresponding water pump model, and the total power of different pump groups after parallel connection is calculated. The one with the lowest total power is selected as the optimal parallel operation plan, thereby performing energy-saving control on the parallel water pumps.

[0132] like Figure 5 As shown in , it describes the change of mean square error during BP neural network training. It can be seen that the mean square error is already ideal after the sixth round of iteration. Figure 6 As shown in Figure 1, it describes the fitness change in the process of optimizing the weight threshold by genetic algorithm in GA-BP neural network. It can be seen that it is already ideal after twenty rounds of iteration.

[0133] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A SHMFF pumping system parallel water pump energy saving control method, characterized in that: The following steps are involved: 11) Acquisition of operating parameters of the parallel water pump group: Acquisition of operating data of the MC100-300 centrifugal water pump and the Multitec A-125 centrifugal water pump in the SHMFF pumping system and preprocessing; 12) Construction of BP neural network: BP neural network A is constructed for MC100-300 centrifugal water pump, and BP neural network B is constructed for Multitec A-125 centrifugal water pump; 13) BP neural network training: BP neural network A and BP neural network B are trained using measured data from MC100-300 centrifugal water pump and Multitec A-125 centrifugal water pump respectively; 14) Optimization of BP neural network A and BP neural network B: Utilize genetic algorithm to optimize the weights and thresholds of BP neural network A and BP neural network B, respectively, to obtain optimized neural network GA-BP A and neural network GA-BP B; The optimization of the BP neural network A and the BP neural network B comprises the following steps: 141) Use genetic algorithm to reduce the fitness function value through multiple iterations of selection, crossover, and mutation operations, thereby continuously optimizing the weights and thresholds of BP neural network A until the allowable mean square error is met: 1411) Population initialization: The individual coding method is real number coding, and the individual coding length is 18+7=25; 1412) Fitness function calculation: After training the BP neural network based on each set of water pump training data, the shaft power output of the system is predicted. The absolute value of the error between the predicted output and the expected output is used as the individual fitness Q. The expected output is the actual measured shaft power value. The calculation formula is as follows: Where N is the number of network output nodes, Y k is the expected output of the kth node of the BP neural network, O k is the predicted output of the kth node, δ is the coefficient, and abs is the absolute value; 1413) Selection operation: Roulette method is used, and the selection strategy is based on the fitness ratio. The selection probability p of each individual i is i for Where Q i and Q j is the fitness value of the individual, δ is the coefficient, and S is the number of individuals in the population; 1414) Crossover operation: Using real number crossover method, the mth chromosome a m and the nth chromosome a n The crossover operation method at position j is as follows: a mj =a mj (1-b)+a nj b a nj =a nj (1-b)+a mj b Where a mj Refers to chromosome a m The jth position, a nj Refers to chromosome a n The jth bit of , b is a random number between [0,1]; 1415) Mutation operation: select the jth gene a of the i-th individual ij Perform mutation, the mutation algorithm is as follows: Where a max Gene a ij The upper bound of a min Gene a ij The next term, H(g)=r2(1-g / G max ) 2 , r2 is a random number, g is the current number of iterations, G max is the maximum number of evolutions, r is a random number between [0,1]; 1416) Update weights and thresholds: After multiple iterations of the genetic algorithm, the optimal individual with the lowest fitness function is obtained, and the weights and thresholds of the BP neural network are assigned to the optimization to obtain the neural network GA-BP A; 142) According to the above method and steps, the weights and thresholds of the BP neural network B are optimized using a genetic algorithm, and after optimization, a neural network GA-BP B is obtained; 15) Generation of the optimal flow distribution scheme: Use the particle swarm optimization method to connect the neural network GA-BP A and the neural network GA-BP B to search for the optimal flow distribution scheme for each pump combination under given demand; The generation of the optimal flow distribution solution includes the following steps: 151) Under a given water supply demand, the particle swarm optimization method is used to connect the neural network GA-BP A and the neural network GA-BP A, the neural network GA-BP A and the neural network GA-BP B, and the neural network GA-BP B and the neural network GA-BP B; Using the single pump head and flow rate (H, Q) vector group as the particle, the particle fitness function F is as follows: Where L is the number of water pumps, L≤m+n, m is the total number of water pumps A, n is the total number of water pumps B, N shaft is the shaft power value of a single pump; 152) Set the number of particles to the number of running pumps, the lift between different particles is the same, and the sum of the flow rates is the required flow rate; 153) The water pump combinations are set to be several MC100-300 centrifugal water pumps connected in parallel, several MultitecA-125 centrifugal water pumps connected in parallel, and several MC100-300 centrifugal water pumps and several MultitecA-125 centrifugal water pumps connected in parallel, and the number of parallel operation units is less than the total number of SHNFF pumping systems; 154) Using the updated particle velocity and position, search for the optimal flow distribution for each pump combination so that the fitness function under the combination is the lowest and generate the optimal flow distribution plan; 16) Energy-saving control of parallel water pumps: The optimal flow distribution scheme, i.e., the head of a single pump and the searched optimal flow distribution, is input into the neural network GA-BP A or neural network GA-BP B corresponding to the corresponding water pump model. The total power of different pump groups after parallel connection is calculated, and the one with the lowest total power is selected as the optimal parallel operation scheme to achieve energy-saving control of parallel water pumps.

2. The energy-saving control method for parallel water pumps in a SHMFF pumping system according to claim 1 is characterized in that: The acquisition of the operating parameters of the parallel water pump group includes the following steps: 21) The head, flow rate, and shaft power values ​​of the working point during single-pump and multi-pump variable frequency operation are measured using the pressure gauges, flow meters, and torque and speed sensors on the inlet and outlet of the MC100-300 centrifugal water pump group, respectively. These values ​​are recorded as Dataset A. 22) The head, flow rate, and shaft power values ​​of the working point when the single pump and multiple pumps are operated with variable frequency are measured using the pressure gauges, flow meters, and torque and speed sensors on the pump shaft of the Multitec A-125 centrifugal water pump group, respectively. These values ​​are recorded as Dataset B. 23) The head, flow rate and shaft power data measured above are normalized and preprocessed. The data normalization formula is as follows: x k =(x k -x min ) / (x max -x min ), Among them, x k is the head, flow or shaft power value, x min is the minimum value of the corresponding data set, x max is the maximum value of the corresponding data set.

3. The energy-saving control method for parallel water pumps in a SHMFF pumping system according to claim 1 is characterized in that: The construction of the BP neural network comprises the following steps: 31) Establish BP neural network A for MC100-300 centrifugal water pump; 311) Assume that BP neural network A includes an input layer, a hidden layer, and an output layer; wherein the input layer has two nodes, representing the head and flow of a single pump, respectively; the output layer has one node, representing the shaft power value of a single pump; the number of hidden layers is calculated according to the formula 2n+1, where n is the number of input nodes; 312) The hidden layer is set to use the tangent S-type transfer function, and the input of the hidden layer neurons is the weighted sum of all inputs, that is, x j =∑ i w ij X i , Output x′ of hidden layer neurons j Use S function to excite x j ,have to but Among them, x j is the input of the hidden layer neurons, w ij is the connection weight between the input layer and the hidden layer, x′ j is the output of the hidden layer neurons; 313) Set the output layer to use a linear transfer function, and the output layer neuron Y k The output is: Among them, the error between the network k-th output and the corresponding ideal output is e k , the error performance index function of the pth sample is N is the number of network output layers; 32) Establish BP neural network B for Multitec A-125 centrifugal water pump.

4. The energy-saving control method for parallel water pumps in a SHMFF pumping system according to claim 1, characterized in that: The training of the BP neural network comprises the following steps: 41) Using the obtained dataset A, 2 / 3 of the data is used to train the neural network A, and the remaining 1 / 3 of the data is used as a validation set for the BP neural network A; 42) Using the obtained dataset B, 2 / 3 of the data is used to train the neural network B, and the remaining 1 / 3 of the data is used as a validation set for the BP neural network B; 43) The training function is set to adopt the Levenberg-Marquardt algorithm, and the network learning function is set to adopt the BP learning rule with momentum term.

Images