Multi-mode adaptive scheduling method for pumping station based on water quantity prediction and dynamic optimization

Abstract

This invention relates to the fields of industrial automation control and smart water management, and discloses a multi-mode adaptive scheduling method for pump stations based on water volume prediction and dynamic optimization. The method includes: calculating the predicted water demand using an LSTM water volume prediction module combined with real-time meteorological characteristics; generating an optimal scheduling instruction set through a particle swarm optimization module with the goal of minimizing total system energy consumption; parsing the instructions and activating the corresponding control mode at the field control terminal, driving the pumps using a strategy combining feedforward compensation and PID control; and adaptively correcting the optimization model parameters using actual operating deviations using a closed-loop feedback analysis module. The system includes a water supply pump unit, a pipeline operation data sensing unit, a field control terminal, and an edge computing decision-making unit. This invention solves the problems of traditional control lag and model distortion, achieving global energy consumption optimization and high-precision stable control of the pump station.

Description

Technical Field

[0001] This invention relates to the fields of industrial automation control and smart water management, specifically to a multi-mode adaptive scheduling method for pump stations based on water volume prediction and dynamic optimization. Background Technology

[0002] As a core unit of municipal infrastructure, urban water supply pumping stations account for a significant portion of water utilities' operating costs. Currently, most water supply pumping stations primarily employ constant pressure variable frequency control, which uses a PID controller to adjust the pump frequency based on the deviation between the real-time detected pressure of the main outlet pipe and the set value to maintain constant network pressure. While this control method meets water supply demands to some extent, the following technical issues still exist in actual operation:

[0003] First, the existing constant pressure control logic is a passive response control, lacking the ability to proactively perceive the water load of the pipeline network. The system cannot predict the trend of water demand changes in the near future and can only make lagging adjustments based on current pressure fluctuations. This lag makes it difficult to coordinate the scheduling strategy at the global level. Pump units often cannot be pre-adjusted to the efficient operating range in future scheduling cycles, easily leading to "over-powered" or frequent start-stop phenomena, resulting in high overall system energy consumption.

[0004] Secondly, in the underlying control execution stage, traditional control systems typically rely entirely on error feedback control. When water load fluctuates significantly and randomly, simply relying on the PID feedback algorithm to eliminate the error requires a certain adjustment time, resulting in a significant lag in system response. During the adjustment process, the pipeline pressure is prone to large overshoot or oscillations, making it difficult to quickly maintain pressure stability when the load changes drastically.

[0005] Third, some pump station control systems employing optimization algorithms typically build their optimization models based on equipment factory parameters or initial installation test data, and these models remain fixed during operation. However, over long-term operation, factors such as impeller wear, decreased motor efficiency, and changes in pipeline resistance coefficients can cause deviations between the actual physical characteristics of the equipment and the initial mathematical model. Due to the lack of closed-loop feedback based on actual operating data and online correction mechanisms for model parameters, the scheduling schemes calculated based on fixed models gradually lose accuracy over time, leading to the failure of the optimization effect. Summary of the Invention

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution: a multi-mode adaptive scheduling method for pumping stations based on water volume prediction and dynamic optimization, applied to a multi-mode adaptive scheduling system for pumping stations based on water volume prediction and dynamic optimization, wherein the system includes a water supply pump unit, a pipeline network operation data sensing unit, a field control terminal, and an edge computing decision unit. The edge computing decision unit is equipped with a data cleaning and normalization module, an LSTM water volume prediction module, a particle swarm dynamic optimization module, and a closed-loop feedback analysis module. The field control terminal stores control logic programs corresponding to constant pressure control mode logic, constant current control mode logic and variable load control mode logic respectively. The method includes the following steps: Step S1, Obtain water demand forecast data: The edge computing decision unit reads historical pipeline operation data and real-time meteorological characteristic data through the industrial bus communication interface and external meteorological data interface; the data cleaning and normalization module performs normalization processing on the data to generate a standardized time series input vector; the LSTM water demand forecast module inputs the time series input vector into a preset long short-term memory neural network model to calculate the water demand forecast value for the next scheduling cycle. Step S2, generating the optimal scheduling scheme: The particle swarm dynamic optimization module receives the predicted water demand value and substitutes it into the preset multi-objective optimization model of the pumping station; The particle swarm dynamic optimization module takes minimizing the total energy consumption of the pumping station system as the objective function and uses the stable pressure range of the pipeline network, the operating frequency range of the pumps, and the start-stop interval of the pumps as constraints. It runs an improved particle swarm algorithm to solve the problem iteratively and outputs an optimal scheduling instruction set that includes the recommended number of pumps in operation, the setpoint of the speed of each pump, and the recommended control mode. Step S3, execute multi-mode adaptive control: The field control terminal receives the optimal scheduling instruction set and parses the recommended control mode therein; the field control terminal activates the corresponding constant pressure control mode logic, constant flow control mode logic, or variable load control mode logic according to the parsing result; the field control terminal uses the water pump speed setpoint in the optimal scheduling instruction set as the base frequency, and superimposes the adjustment frequency obtained by PID calculation of the difference between the actual pipeline pressure data and the target setpoint corresponding to the current control mode onto the base frequency to generate the final frequency control signal and start / stop control signal, driving the water supply pump unit to operate; Step S4, closed-loop feedback and parameter correction: The pipeline network operation data sensing unit collects actual pipeline network operation status data within the current scheduling cycle; the closed-loop feedback analysis module calculates the deviation value between the actual operation data and the optimization target in step S2, and transmits the deviation value to the particle swarm dynamic optimization module; The particle swarm dynamic optimization module dynamically adjusts the penalty term coefficients or power fitting formula coefficients in the multi-objective optimization model of the pump station according to the magnitude of the deviation value, and applies the adjusted coefficients in the optimization calculation of the next scheduling cycle.

[0007] Preferably, in step S1, the LSTM water volume prediction module is internally configured with a two-layer stacked neural network structure, which includes an input layer, a first long short-term memory layer, a second long short-term memory layer, a regularized dropout layer, and a fully connected output layer. The input layer receives a three-dimensional input tensor containing batch size, time step, and feature dimension; the fully connected output layer outputs a normalized prediction result value and performs inverse normalization processing to obtain the water demand prediction value. The specific process of the inverse normalization process is as follows: multiply the normalized prediction result by the difference between the historical maximum and historical minimum values ​​of the flow characteristics in the training dataset, and then add the historical minimum value to the product result to restore the predicted water demand value.

[0008] Preferably, in step S2, the objective function of the multi-objective optimization model of the pumping station is configured to calculate the total operating cost of the system, which consists of an energy consumption cost component and a start-up and shutdown loss component. The energy cost component is obtained by accumulating the product of the shaft power of all operating water pumps in each time period within the rolling optimization window and the length of the time period; The start-stop loss component is obtained by accumulating the product of the number of start-stop state switching times of all water pumps within the rolling optimization window and the preset start-stop penalty coefficient; The objective function takes the sum of the energy consumption cost component and the start-stop loss component as the optimization objective.

[0009] Preferably, in step S2, the improved particle swarm algorithm uses a nonlinear dynamic reduction strategy to adjust the inertia weight during the iteration process; The specific process of the nonlinear dynamic reduction strategy is as follows: calculate the weight decay based on the square of the ratio of the current iteration number to the maximum iteration number; Multiply the difference between the preset maximum and preset minimum values ​​of the inertia weight by the weight decay amount, and then subtract the product from the preset maximum value to obtain the inertia weight for the current iteration.

[0010] Preferably, in step S2, the improved particle swarm algorithm further includes a chaotic perturbation mechanism; When the global optimal position has not been updated in a preset number of consecutive iterations, the global optimal position is mutated using a Logistic mapping. The specific process of the Logistic mapping is as follows: multiply the current normalized particle position mapping variable by the chaotic control parameter, and then multiply it by the difference between the current normalized particle position mapping variable and the original value to obtain the new mapping variable after chaotic iteration.

[0011] Preferably, the field control terminal extracts the recommended frequency setting value contained in the optimal scheduling instruction set and defines it as the feedforward base frequency; The field control terminal runs an incremental PID algorithm to calculate the pressure deviation between the actual pressure data of the pipeline network and the target set value, and calculates the output feedback adjustment frequency increment based on the proportional coefficient, integral coefficient and derivative coefficient. The field control terminal uses integral accumulation logic to superimpose the feedforward base frequency with the historical feedback adjustment frequency to generate the final inverter drive command.

[0012] Preferably, the specific process of the integral accumulation logic is as follows: Calculate the cumulative sum of all PID feedback adjustment frequency increments from the start of control to the current time; The feedforward base frequency is added to the accumulated sum to obtain the final drive frequency command issued to the frequency converter at the current moment.

[0013] Preferably, in step S4, the deviation value includes flow prediction deviation and energy consumption estimation deviation; The closed-loop feedback analysis module feeds back the flow prediction deviation to the LSTM water prediction module to correct the weight parameters of the neural network. The closed-loop feedback analysis module feeds the energy consumption estimation deviation back to the particle swarm dynamic optimization module, which is used to calibrate the pump power fitting formula coefficients in the pump station multi-objective optimization model. When the actual energy consumption is higher than the theoretical energy consumption and the deviation exceeds the preset threshold, the coefficient of the constant term in the power fitting formula is increased to achieve adaptive aging compensation of the model parameters.

[0014] Preferably, a multi-mode adaptive scheduling system for pumping stations based on water volume prediction and dynamic optimization, the system comprising: The water supply pump unit is located inside the pumping station and is equipped with a first variable frequency water pump, a second variable frequency water pump and a third fixed frequency water pump. The pipeline operation data sensing unit is set at key nodes of the pump station's outlet pipeline network and is equipped with a pressure sensor for collecting actual pressure data of the pipeline network and an electromagnetic flow meter for collecting instantaneous flow data of the pipeline network. The field control terminal is equipped with analog input modules, digital output modules, analog output modules, and an industrial bus communication interface; The digital output module is equipped with an intermediate relay and is connected to the power frequency start controller of the third power frequency water pump through the intermediate relay. The field control terminal stores control logic programs corresponding to constant pressure control mode logic, constant current control mode logic and variable load control mode logic respectively. The edge computing decision unit establishes a bidirectional data transmission channel with the field control terminal through the industrial bus communication interface; The edge computing decision unit integrates a processor and a memory. The memory stores computer instructions that, when executed by the processor, implement the data cleaning and normalization module, the LSTM water volume prediction module, the particle swarm dynamic optimization module, and the closed-loop feedback analysis module. The system is used to execute the above-mentioned multi-mode adaptive scheduling method for pumping stations based on water volume prediction and dynamic optimization.

[0015] Preferably, a computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the above method.

[0016] This invention provides a multi-mode adaptive scheduling method for pumping stations based on water volume prediction and dynamic optimization. It has the following beneficial effects: (1) This invention achieves global energy consumption optimization for pump station operation through the collaborative work of the LSTM water volume prediction module and the particle swarm dynamic optimization module within the edge computing decision unit. The LSTM water volume prediction module accurately calculates the predicted water demand for the next scheduling cycle using historical pipeline operation data and real-time meteorological characteristic data, providing a forward-looking basis for scheduling; the particle swarm dynamic optimization module further substitutes this predicted value into a model with the objective function of minimizing the total energy consumption of the pump station system. This prediction-based rolling optimization mechanism breaks the limitation of traditional control relying solely on current pressure feedback, ensuring that the water supply pump unit always operates within the optimal efficiency range, effectively reducing the overall energy consumption of the system.

[0017] (2) This invention employs a control strategy combining feedforward compensation and feedback regulation at the execution level of the field control terminal, significantly improving the system's response speed and stability. The field control terminal does not rely solely on the PID algorithm to eliminate errors; instead, it directly defines the pump speed setpoint in the optimal scheduling instruction set as the feedforward base frequency, and only superimposes the adjustment frequency obtained by PID calculation based on the difference between the actual pipeline pressure data and the target setpoint onto the base frequency. This method utilizes the theoretically optimal operating point calculated through optimization as the control benchmark, significantly reducing the burden and regulation lag of the PID controller, and avoiding large fluctuations or overshoot in the pipeline pressure.

[0018] (3) This invention utilizes a closed-loop feedback analysis module to construct an adaptive correction mechanism for model parameters, solving the model distortion problem caused by equipment aging. The system can calculate the deviation between actual operating data and the optimization objective, and feed back the energy consumption estimation deviation to the particle swarm dynamic optimization module. By dynamically adjusting the penalty term coefficient or power fitting formula coefficient in the multi-objective optimization model of the pump station according to the magnitude of the deviation, the system can automatically compensate for the impact of pump performance degradation, ensuring that the scheduling model maintains high-precision mapping capability during long-term operation and maintaining the effectiveness of scheduling decisions. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the system framework of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 2 This invention provides a multi-mode adaptive scheduling system for pumping stations based on water volume prediction and dynamic optimization, comprising: a water supply pump unit, a pipeline network operation data sensing unit, a field control terminal, and an edge computing decision-making unit. Specifically: The water supply pump unit is located inside the pumping station and is used to perform pressurized water supply tasks. The water supply pump unit includes a first variable frequency pump, a second variable frequency pump, and a third fixed frequency pump. The inlet ends of the first variable frequency pump, the second variable frequency pump, and the third fixed frequency pump are connected to the municipal water supply network, and the outlet ends converge to the main outlet pipe of the pumping station. The motor input end of the first variable frequency pump is electrically connected to the output end of the first variable frequency drive, the motor input end of the second variable frequency pump is electrically connected to the output end of the second variable frequency drive, and the motor input end of the third fixed frequency pump is electrically connected to the output end of the fixed frequency start controller.

[0022] The first and second frequency converters have frequency adjustment interfaces, which are used to receive external frequency commands and adjust the speed of the corresponding water pump motor. The power frequency start controller uses a circuit breaker or a soft starter to control the start and stop status of the third power frequency water pump.

[0023] The pipeline operation data sensing unit is set at key nodes of the pump station's outlet pipeline network. The pipeline operation data sensing unit includes a pressure sensor and an electromagnetic flow meter. The pressure sensor is installed at the pump station's outlet main pipe or the most unfavorable point of the pipeline network to collect real-time actual pressure data of the pipeline network. The electromagnetic flow meter is installed on the straight section of the outlet main pipe to collect real-time instantaneous flow data of the pipeline network.

[0024] The field control terminal adopts a programmable logic controller and includes an analog input module, a digital output module, an analog output module, and an industrial bus communication interface. The digital output module is connected to the control terminal of the power frequency start controller through an intermediate relay.

[0025] The analog input module is connected to the pressure sensor and electromagnetic flow meter via signal transmission cables to receive actual pressure data and instantaneous flow data of the pipeline network. The digital output module is connected to the control terminal of the power frequency start controller to output start and stop control signals. The analog output module is connected to the control terminals of the first and second frequency converters to output frequency adjustment signals.

[0026] The edge computing decision unit adopts an industrial computer or edge server. The edge computing decision unit establishes a bidirectional data transmission channel with the field control terminal through an industrial bus communication interface. This bidirectional data transmission channel is based on the Ethernet or serial bus physical layer and runs the OPC UA communication protocol or Modbus TCP communication protocol.

[0027] The edge computing decision unit integrates a data storage device and a processor. The data storage device stores computer program instructions, which, when executed by the processor, can realize the following functional modules: data cleaning and normalization module, LSTM water volume prediction module, particle swarm dynamic optimization module, and closed-loop feedback analysis module.

[0028] The data cleaning and normalization module is used to preprocess the received historical operating data, and the LSTM water demand prediction module contains a pre-trained long short-term memory neural network model, which is used to calculate the predicted value of future water demand based on the preprocessed data.

[0029] The particle swarm optimization module includes a multi-objective optimization algorithm, which is used to calculate the optimal scheduling instruction set based on the predicted future water demand. The optimal scheduling instruction set includes pump start / stop combination instructions, target frequency setting instructions, and control mode switching instructions.

[0030] The closed-loop feedback analysis module is used to collect deviation data generated during the actual operation of the system and feed this deviation data back to the particle swarm dynamic optimization module to correct the algorithm parameters.

[0031] The field control terminal stores a control logic program, which includes multiple preset PID control function blocks, corresponding to constant pressure control mode logic, constant current control mode logic, and variable load control mode logic, respectively. The field control terminal selectively activates the constant pressure control mode logic, constant current control mode logic, or variable load control mode logic according to the control mode switching instructions received from the optimal scheduling instruction set.

[0032] The external meteorological data interface communicates with the edge computing decision unit through the network communication module to obtain real-time meteorological feature data. The real-time meteorological feature data is transmitted to the LSTM water volume prediction module as an auxiliary input variable.

[0033] Please see Figure 1 Based on the above, this invention provides a multi-mode adaptive scheduling method for pumping stations based on water volume prediction and dynamic optimization, specifically including the following steps: Step S1: Obtain water demand forecast data. The edge computing decision unit periodically reads historical pipeline operation data and real-time meteorological characteristic data through the industrial bus communication interface and external meteorological data interface. The data cleaning and normalization module performs outlier removal and maximum-minimum normalization processing on the above data to generate a standardized time series input vector. The LSTM water demand forecast module inputs the time series input vector into the preset long short-term memory neural network model. Through the forget gate, input gate and output gate operations inside the model, the water demand forecast value for the next scheduling cycle is calculated.

[0034] Step S2: Generate the optimal scheduling scheme. The particle swarm dynamic optimization module receives the water demand prediction value and substitutes it into the preset multi-objective optimization model of the pumping station. The particle swarm dynamic optimization module takes minimizing the total energy consumption of the pumping station system as the objective function and the stable pressure range of the pipeline network, the pump operating frequency range, and the pump start-stop interval as constraints. It runs the improved particle swarm algorithm for iterative solution. The solution process is based on the rolling window mechanism, triggering a calculation every preset time step, and outputs the optimal scheduling instruction set including the recommended number of pumps in operation, the set value of the speed of each pump, and the recommended control mode.

[0035] Step S3: Perform multi-mode adaptive control. The field control terminal receives the optimal scheduling instruction set and parses the recommended control mode. Based on the parsing result, the field control terminal activates the corresponding constant pressure control mode logic, constant current control mode logic, or variable load control mode logic. The field control terminal uses the pump speed setpoint in the optimal scheduling instruction set as the base frequency, and superimposes the adjustment frequency obtained by PID calculation of the difference between the actual pipeline pressure data and the target setpoint onto the base frequency to generate the final frequency control signal and start / stop control signal, driving the first frequency converter, the second frequency converter, and the power frequency start controller to operate.

[0036] Step S4, closed-loop feedback and parameter correction: The pipeline operation data sensing unit continuously collects actual pipeline operation status data within the current scheduling cycle. The closed-loop feedback analysis module calculates the deviation between the actual operation data and the optimization target in step S2. The closed-loop feedback analysis module transmits the deviation value to the particle swarm dynamic optimization module. The particle swarm dynamic optimization module dynamically adjusts the penalty term coefficient in the pump station energy consumption model according to the magnitude of the deviation value and applies the adjusted coefficient in the optimization calculation of the next scheduling cycle.

[0037] Furthermore, this method is executed by the data cleaning and normalization module and the LSTM water volume prediction module within the edge computing decision unit.

[0038] The data cleaning and normalization module establishes a multidimensional time series feature matrix. The feature dimensions of this multidimensional time series feature matrix include: historical pipeline flow data, historical pipeline pressure data, and real-time meteorological feature data. The data cleaning and normalization module uses the max-min normalization algorithm to map the values ​​in the multidimensional time series feature matrix to the range of zero to one.

[0039] The specific calculation logic for normalization follows the formula:

[0040] in, This represents the normalized input value, used for computation in the neural network. This represents the original collected observation data, including specific flow rate, pressure, or meteorological values; This represents the minimum value of the feature within the historical sliding window; This represents the maximum value of the feature within the historical sliding window.

[0041] The data cleaning and normalization module constructs a three-dimensional input tensor from the normalized data. The shape of this three-dimensional input tensor is defined as ( , , ),in, Represents batch size; The time step is set to 24, corresponding to historical data from the past two hours. The representative feature dimension is set to 3, corresponding to flow rate, pressure, and meteorological features respectively.

[0042] The LSTM water volume prediction module is internally configured with a two-layer stacked neural network structure, which includes an input layer, a first long short-term memory layer, a second long short-term memory layer, a regularized dropout layer, and a fully connected output layer.

[0043] The input layer receives the three-dimensional input tensor output by the data cleaning and normalization module. The first long short-term memory layer contains 64 long short-term memory neural units. The first long short-term memory layer uses a pre-trained weight matrix to extract temporal features from the three-dimensional input tensor and outputs the first layer hidden state vector.

[0044] The second long short-term memory layer contains 32 long short-term memory neural units. The second long short-term memory layer is cascaded with the first long short-term memory layer. It receives the hidden state vector of the first layer and performs a high-order nonlinear transformation to output the hidden state vector of the second layer.

[0045] The regularized dropout layer is set after the second long short-term memory layer. The regularized dropout layer is configured with a dropout rate parameter of 0.2. This parameter is used to randomly suppress neuron activation to prevent overfitting during the offline training phase of the model. During the online prediction phase, this layer is in a pass-through state.

[0046] The fully connected output layer is connected to the regularized dropout layer. The fully connected output layer receives the hidden state vector from the second layer and maps it to the normalized prediction result value. The LSTM water quantity prediction module provides normalized prediction results. After performing inverse normalization, the final water demand forecast (denoted as ) is calculated. ).

[0047] The specific calculation logic for inverse normalization follows the formula:

[0048] in, This represents the actual water demand prediction value (unit: cubic meters per hour) output to the optimization module. The normalized value (between 0 and 1) represents the output of the fully connected output layer. This represents the historical maximum value of the traffic feature in the training dataset; This represents the historical minimum value of the traffic feature in the training dataset.

[0049] Before being put into operation, the LSTM water prediction module was pre-trained offline. During the offline training phase, the root mean square error was used as the loss function. The Adam optimization algorithm was used to backpropagate the gradient of the loss function to iteratively update the weight matrix and bias term of the first and second long short-term memory layers until the loss function converged to the preset threshold. During the online operation phase, the LSTM water prediction module kept the weight matrix and bias term fixed and only performed forward propagation calculations.

[0050] In step S2 above, the particle swarm dynamic optimization module generates an optimal scheduling instruction set based on a preset mathematical model and solution algorithm. The specific construction and execution logic of the mathematical model and solution algorithm are as follows: First, the particle swarm optimization module constructs a multi-objective optimization model for pump station operation. This model is defined as minimizing the weighted sum of the total system power consumption and the pump start-up / stop penalty values ​​within the scheduling cycle. The mathematical expression for the objective function of the multi-objective optimization model is as follows:

[0051] in, The calculated value of the objective function is used to characterize the quality of the current scheduling scheme; This represents the total number of time periods within the scrolling optimization window (for example, setting it to 4 represents four 15-minute time periods within the next hour). This represents the total number of water pumps in the water supply pump unit. Representing the The water pump is in the first Shaft power for each time period (unit: kilowatt); Represents the duration of a single scheduling period (unit: hours); This represents the start-stop penalty coefficient (unit: kilowatt-hours / time), which is used to quantify the impact of pump start-stop actions on equipment lifespan. Representing the The number of times the pump switches between start and stop states within the entire optimization window (i.e., the total number of state changes from running to stopping or from stopping to running).

[0052] To calculate the objective function, the particle swarm optimization module is equipped with pump characteristic calculation logic, which calculates the shaft power based on the pump's physical characteristic curve. The calculation process is based on the flow-efficiency curve and flow-head curve pre-stored in the memory, mapping the real-time operating frequency and flow of the water pump into shaft power data.

[0053] In addition, the multi-objective optimization model also includes a set of constraints to limit the solution space. The set of constraints includes flow balance constraints, pressure stability constraints, and equipment operation constraints. The mathematical expression for the flow balance constraint is as follows:

[0054] in, Represents the first input from the LSTM water prediction module Water demand forecast for each time period; Representing the The water pump is in the first Output flow rate for each time period; This represents the allowable flow deviation ratio, which is set to 0.05 in this embodiment.

[0055] Pressure stability constraints limit the pressure at pipeline nodes to be maintained within a preset target pressure range. Equipment operation constraints limit the operating frequency of variable frequency pumps to be between the minimum safe frequency and the power frequency, and limit the number of pumps operating simultaneously to not exceed the maximum number of pumps operating in parallel.

[0056] For the above multi-objective optimization model, the particle swarm dynamic optimization module runs an improved particle swarm optimization algorithm to solve the problem. The algorithm initializes a particle population, where each particle represents a potential pump station scheduling scheme. The scheduling scheme includes the combination of pump start and stop states and the frequency setting value of the variable frequency pump.

[0057] The improved particle swarm optimization algorithm searches for the global optimum by iteratively updating the velocity and position of particles. The velocity update formula is as follows:

[0058] The position update formula is as follows:

[0059] in, Representing the In the nth iteration A particle velocity vector of dimension; Representing the In the nth iteration A dimensional particle position vector (corresponding to a specific frequency value or start / stop state). The inertia weight represents the current iteration; Represents self-awareness learning factors; Represents social cognitive learning factors; and Represents a uniformly distributed random number within the interval [0,1]. The historical best position representing an individual particle; This represents the global optimal position of the population.

[0060] To avoid the algorithm getting trapped in local optima, the particle swarm optimization module uses a nonlinear dynamic reduction strategy to adjust the inertia weights during the iteration process. , Inertia weight The calculation formula is as follows:

[0061] in, The preset maximum value represents the inertia weight; The preset minimum value representing the inertia weight; The index represents the current iteration count; This represents the preset maximum number of iterations.

[0062] The particle swarm optimization module also employs an adaptive strategy to adjust the learning factor. It decreases linearly with the number of iterations. It increases linearly with the number of iterations.

[0063] Furthermore, when the global optimal position If no update occurs during a preset number of iterations, the particle swarm optimization module triggers a chaotic perturbation mechanism, which uses a Logistic mapping to mutate the global optimal position.

[0064] The formula for calculating the Logistic mapping is as follows:

[0065] in, Represents the normalized particle position mapping variable; This represents the chaos control parameter, with a value of 4 to ensure that the system is in a completely chaotic state. This represents the mapped variable after one chaotic iteration, used to make small perturbations to the original particle's position.

[0066] The particle swarm optimization module reaches its maximum number of iterations. Finally, output the final globally optimal position. The globally optimal position is decoded into the optimal scheduling instruction set, which specifies the pump number to be started in the next scheduling cycle, the target frequency setting value for each variable frequency pump, and the recommended system control mode.

[0067] Furthermore, the field control terminal receives the optimal scheduling instruction set from the edge computing decision unit through the industrial bus communication interface. The instruction parsing logic inside the field control terminal first reads the control mode identifier bit in the optimal scheduling instruction set. If the control mode identifier bit indicates constant voltage control, the field control terminal activates the constant voltage control mode logic; if it indicates constant current control, it activates the constant current control mode logic; if it indicates variable load adaptive control, it activates the variable load control mode logic.

[0068] When the constant voltage control mode logic is activated, the field control terminal executes a composite control strategy combining feedforward compensation and feedback regulation. The field control terminal extracts the recommended frequency setpoint contained in the optimal scheduling instruction set and defines it as the feedforward base frequency. The feedforward fundamental frequency This corresponds to the theoretically optimal operating point calculated by the multi-objective dynamic optimization model.

[0069] Meanwhile, the field control terminal acquires the actual pressure data of the pipeline network through the analog input module and compares it with the preset target pressure value to calculate the pressure deviation. The field control terminal uses an incremental PID algorithm to monitor pressure deviations. Perform calculations and output feedback to adjust the frequency increment. .

[0070] The operational logic of the incremental PID algorithm follows the mathematical relationship below:

[0071] in, Represents the current control cycle The calculated frequency increment adjustment value; Represents the proportionality coefficient; Represents the integral coefficient; Represents the differential coefficient; This represents the difference between the target pressure and the actual pressure of the pipeline at the current sampling moment. Represents the previous sampling time Pressure deviation value; Represents the upsampling time The pressure deviation value.

[0072] The field control terminal uses integral accumulation logic to feedforward the base frequency. The frequency is superimposed with the historical feedback adjustment frequency to generate the final inverter drive command. .

[0073] The superposition calculation formula is:

[0074] in, Representing the The frequency command (unit: Hertz) that is ultimately sent to the hardware at any given moment. This represents the reference frequency calculated and issued by the particle swarm dynamic optimization module, serving as the feedforward control quantity. Represents the moment control begins Up to the current moment The sum of all PID frequency increment adjustment values, i.e. the integral output term of the PID controller, is used to eliminate system steady-state error.

[0075] The field control terminal transmits the inverter drive commands through the analog output module. It is converted into a 4-20mA analog current signal or a 0-10V analog voltage signal and sent to the first frequency converter and the second frequency converter.

[0076] For the pump start / stop commands in the optimal scheduling command set, the field control terminal outputs high and low level signals through the digital output module. These high and low level signals drive the intermediate relay to operate, thereby controlling the power frequency start controller to close or open, thus enabling the third power frequency pump to be put into operation or cut off.

[0077] After completing the control execution of a scheduling cycle, the system enters the closed-loop feedback correction stage in step S4. The pipeline operation data sensing unit records the average actual flow rate, average actual pressure, and total energy consumption of the pumping station within the scheduling cycle.

[0078] The closed-loop feedback analysis module within the edge computing decision unit acquires the aforementioned actual operational data and compares it with the predicted water demand at the start of the scheduling cycle and the theoretical energy consumption value of the multi-objective optimization model. It then calculates the prediction deviation vector, which includes the flow prediction deviation. and energy consumption estimation deviation .

[0079] The closed-loop feedback analysis module will analyze the traffic prediction deviation. The error is fed back to the LSTM water prediction module, which stores the deviation in the historical error database and increases the weight of the high-error sample during the next offline training update to correct the weight parameters of the neural network.

[0080] The closed-loop feedback analysis module will analyze the energy consumption estimation deviation. Feedback is sent to the particle swarm optimization module, which estimates the deviation based on energy consumption. The sign and amplitude of the coefficients are used to calibrate the coefficients of the pump power fitting formula in the multi-objective optimization model. When the actual energy consumption is higher than the theoretical energy consumption and the deviation exceeds the preset threshold, the particle swarm dynamic optimization module increases the coefficient of the constant term in the power fitting formula so that the subsequent optimization calculation is closer to the actual performance curve of the equipment after aging, thereby realizing adaptive aging compensation of the model parameters.

[0081] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-mode adaptive scheduling method for pumping stations based on water volume prediction and dynamic optimization, characterized in that, An application is made to a multi-mode adaptive scheduling system for pumping stations based on water volume prediction and dynamic optimization. The system includes a water supply pump unit, a pipeline network operation data sensing unit, a field control terminal, and an edge computing decision unit. The edge computing decision unit is equipped with a data cleaning and normalization module, an LSTM water volume prediction module, a particle swarm dynamic optimization module, and a closed-loop feedback analysis module. The field control terminal stores control logic programs corresponding to constant pressure control mode logic, constant current control mode logic and variable load control mode logic respectively. The method includes the following steps: Step S1, Obtain water demand forecast data: The edge computing decision unit reads historical pipeline operation data and real-time meteorological characteristic data through the industrial bus communication interface and external meteorological data interface; the data cleaning and normalization module performs normalization processing on the data to generate a standardized time series input vector; the LSTM water demand forecast module inputs the time series input vector into a preset long short-term memory neural network model to calculate the water demand forecast value for the next scheduling cycle. Step S2, generating the optimal scheduling scheme: The particle swarm dynamic optimization module receives the predicted water demand value and substitutes it into the preset multi-objective optimization model of the pumping station; The particle swarm dynamic optimization module takes minimizing the total energy consumption of the pumping station system as the objective function and uses the stable pressure range of the pipeline network, the operating frequency range of the pumps, and the start-stop interval of the pumps as constraints. It runs an improved particle swarm algorithm to solve the problem iteratively and outputs an optimal scheduling instruction set that includes the recommended number of pumps in operation, the setpoint of the speed of each pump, and the recommended control mode. Step S3, execute multi-mode adaptive control: The field control terminal receives the optimal scheduling instruction set and parses the recommended control mode therein; the field control terminal activates the corresponding constant pressure control mode logic, constant flow control mode logic, or variable load control mode logic according to the parsing result; the field control terminal uses the water pump speed setpoint in the optimal scheduling instruction set as the base frequency, and superimposes the adjustment frequency obtained by PID calculation of the difference between the actual pipeline pressure data and the target setpoint corresponding to the current control mode onto the base frequency to generate the final frequency control signal and start / stop control signal, driving the water supply pump unit to operate; Step S4, closed-loop feedback and parameter correction: The pipeline network operation data sensing unit collects actual pipeline network operation status data within the current scheduling cycle; the closed-loop feedback analysis module calculates the deviation value between the actual operation data and the optimization target in step S2, and transmits the deviation value to the particle swarm dynamic optimization module; The particle swarm dynamic optimization module dynamically adjusts the penalty term coefficients or power fitting formula coefficients in the multi-objective optimization model of the pump station according to the magnitude of the deviation value, and applies the adjusted coefficients in the optimization calculation of the next scheduling cycle.

2. The multi-mode adaptive scheduling method for pumping stations based on water volume prediction and dynamic optimization according to claim 1, characterized in that: In step S1, the LSTM water volume prediction module is internally configured with a two-layer stacked neural network structure, which includes an input layer, a first long short-term memory layer, a second long short-term memory layer, a regularized discard layer, and a fully connected output layer. The input layer receives a three-dimensional input tensor containing batch size, time step, and feature dimension; the fully connected output layer outputs a normalized prediction result value and performs inverse normalization processing to obtain the water demand prediction value. The specific process of the inverse normalization process is as follows: multiply the normalized prediction result by the difference between the historical maximum and historical minimum values ​​of the flow characteristics in the training dataset, and then add the historical minimum value to the product result to restore the predicted water demand value.

3. The multi-mode adaptive scheduling method for pumping stations based on water volume prediction and dynamic optimization according to claim 1, characterized in that: In step S2, the objective function of the multi-objective optimization model of the pump station is configured to calculate the total operating cost of the system, which consists of an energy consumption cost component and a start-up and shutdown loss component. The energy cost component is obtained by accumulating the product of the shaft power of all operating water pumps in each time period within the rolling optimization window and the length of the time period; The start-stop loss component is obtained by accumulating the product of the number of start-stop state switching times of all water pumps within the rolling optimization window and the preset start-stop penalty coefficient; The objective function takes the sum of the energy consumption cost component and the start-stop loss component as the optimization objective.

4. The multi-mode adaptive scheduling method for pumping stations based on water volume prediction and dynamic optimization according to claim 1, characterized in that: In step S2, the improved particle swarm algorithm uses a nonlinear dynamic reduction strategy to adjust the inertia weight during the iteration process. The specific process of the nonlinear dynamic reduction strategy is as follows: calculate the weight decay based on the square of the ratio of the current iteration number to the maximum iteration number; Multiply the difference between the preset maximum and preset minimum values ​​of the inertia weight by the weight decay amount, and then subtract the product from the preset maximum value to obtain the inertia weight for the current iteration.

5. The multi-mode adaptive scheduling method for pumping stations based on water volume prediction and dynamic optimization according to claim 4, characterized in that: In step S2, the improved particle swarm algorithm further includes a chaotic perturbation mechanism; When the global optimal position has not been updated in a preset number of consecutive iterations, the global optimal position is mutated using a Logistic mapping. The specific process of the Logistic mapping is as follows: multiply the current normalized particle position mapping variable by the chaotic control parameter, and then multiply it by the difference between the current normalized particle position mapping variable and the original value to obtain the new mapping variable after chaotic iteration.

6. The multi-mode adaptive scheduling method for pumping stations based on water volume prediction and dynamic optimization according to claim 1, characterized in that: The field control terminal extracts the recommended frequency setting value contained in the optimal scheduling instruction set and defines it as the feedforward base frequency; The field control terminal runs an incremental PID algorithm to calculate the pressure deviation between the actual pressure data of the pipeline network and the target set value, and calculates the output feedback adjustment frequency increment based on the proportional coefficient, integral coefficient and derivative coefficient. The field control terminal uses integral accumulation logic to superimpose the feedforward base frequency with the historical feedback adjustment frequency to generate the final inverter drive command.

7. The multi-mode adaptive scheduling method for pumping stations based on water volume prediction and dynamic optimization according to claim 6, characterized in that: The specific process of the integral accumulation logic is as follows: Calculate the cumulative sum of all PID feedback adjustment frequency increments from the start of control to the current time; The feedforward base frequency is added to the accumulated sum to obtain the final drive frequency command issued to the frequency converter at the current moment.

8. The multi-mode adaptive scheduling method for pumping stations based on water volume prediction and dynamic optimization according to claim 1, characterized in that: In step S4, the deviation value includes flow prediction deviation and energy consumption estimation deviation; The closed-loop feedback analysis module feeds back the flow prediction deviation to the LSTM water prediction module to correct the weight parameters of the neural network. The closed-loop feedback analysis module feeds the energy consumption estimation deviation back to the particle swarm dynamic optimization module, which is used to calibrate the pump power fitting formula coefficients in the pump station multi-objective optimization model. When the actual energy consumption is higher than the theoretical energy consumption and the deviation exceeds the preset threshold, the coefficient of the constant term in the power fitting formula is increased to achieve adaptive aging compensation of the model parameters.

9. A multi-mode adaptive scheduling system for pumping stations based on water volume prediction and dynamic optimization, characterized in that, The system includes: The water supply pump unit is located inside the pumping station and is equipped with a first variable frequency water pump, a second variable frequency water pump and a third fixed frequency water pump. The pipeline operation data sensing unit is set at key nodes of the pump station's outlet pipeline network and is equipped with a pressure sensor for collecting actual pressure data of the pipeline network and an electromagnetic flow meter for collecting instantaneous flow data of the pipeline network. The field control terminal is equipped with analog input modules, digital output modules, analog output modules, and an industrial bus communication interface; The digital output module is equipped with an intermediate relay and is connected to the power frequency start controller of the third power frequency water pump through the intermediate relay. The field control terminal stores control logic programs corresponding to constant pressure control mode logic, constant current control mode logic and variable load control mode logic respectively. The edge computing decision unit establishes a bidirectional data transmission channel with the field control terminal through the industrial bus communication interface; The edge computing decision unit integrates a processor and a memory. The memory stores computer instructions that, when executed by the processor, implement the data cleaning and normalization module, the LSTM water volume prediction module, the particle swarm dynamic optimization module, and the closed-loop feedback analysis module. The system is used to execute the multi-mode adaptive scheduling method for pumping stations based on water volume prediction and dynamic optimization as described in any one of claims 1 to 8.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images