A deep learning-based intelligent water management system

Abstract

This invention relates to the field of smart water management technology, and more particularly to a deep learning-based smart water management system, comprising: a data acquisition and joint prediction module, a health assessment module, a constraint scheduling module, an adaptive correction module, a simulation arbitration module, and an instruction execution module. By acquiring multi-source operational data and jointly predicting liquid level and inflow sequences, the system identifies pump health scores online. The predicted information, electricity price periods, and health scores are input into a deep reinforcement learning scheduling model to generate scheduling actions. After adaptive correction based on health scores and actual characteristic parameters, the system is arbitrated through real-time simulation by a pipeline hydrodynamic simulation model. If safety conditions are met, the final control instruction is output to execute the pump scheduling. This invention achieves energy-saving optimized scheduling of pump stations and equipment health protection under hard constraints of flood control safety, improving system safety and economy, and effectively solving the problems of separation between prediction and scheduling and difficulty in balancing safety and energy saving in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of smart water technology, and in particular to a smart water management system based on deep learning. Background Technology

[0002] Urban drainage pumping stations are critical infrastructure for ensuring urban flood control and drainage safety. Their operation and scheduling must simultaneously address the dual pressures of flood control safety and energy conservation. Regarding flood control, if the water level in the pumping station's forebay is too high during heavy rains, it can easily lead to secondary disasters such as road flooding and sewage overflow, severely impacting urban operations and residents' lives. In terms of energy conservation, as high-power electrical facilities, pumping stations incur significant electricity costs, which constitute a substantial proportion of water utility operating costs. Furthermore, the implementation of peak-valley time-of-use pricing policies results in substantial differences in electricity costs at different times. Therefore, pumping station scheduling is essentially a multi-objective decision-making problem seeking a dynamic balance between "maintaining low water levels to ensure flood control capacity" and "utilizing off-peak hours for drainage to reduce electricity costs." In recent years, with the widespread adoption of IoT sensing technology, SCADA systems, and meteorological forecast data interfaces, pumping stations have accumulated massive amounts of real-time operating data, rainfall forecast data, and electricity price time-series data, providing a foundation for data-driven intelligent scheduling methods.

[0003] However, traditional scheduling strategies typically rely on fixed rules set by human experience. For example, reducing the number of pumps running during peak electricity price periods to "avoid peak" operations, or forcibly draining water when the water level exceeds the warning value to prioritize "flood control." However, these rules lack comprehensive prediction of future rainfall trends, pipeline storage capacity, and equipment health status. They cannot proactively drain water before heavy rain to free up storage capacity, nor can they maximize the use of off-peak electricity prices for economical drainage while ensuring flood control safety. Although some existing studies have introduced water level or flow prediction models, the prediction results are disconnected from scheduling decisions. They fail to coordinate and optimize rainfall forecasts, time-of-use electricity prices, pipeline hydraulic characteristics, and real-time pump operating conditions. As a result, scheduling strategies struggle to simultaneously meet flood control safety constraints and energy conservation requirements when facing complex operating conditions with strong temporal coupling and multiple conflicting objectives. Summary of the Invention

[0004] To address the above shortcomings, this invention provides a deep learning-based intelligent water management system, aiming to improve the existing pump station scheduling methods' lack of comprehensive prediction and coordinated optimization of rainfall forecasts, time-of-use electricity pricing, and pipeline storage capacity, making it difficult to achieve energy-saving and consumption-reducing operation of pump stations while meeting the hard constraints of flood control safety.

[0005] This invention provides the following technical solution: a smart water management system based on deep learning, comprising:

[0006] The data acquisition and joint prediction module is used to acquire multi-source operation data of the pumping station and generate liquid level prediction sequence and inflow rate prediction sequence for a future preset time period through a joint time series prediction model.

[0007] The health identification module is used to identify the actual characteristic parameters of each water pump online through the water pump electrical parameter data of the multi-source operation data, and determine the health score of each water pump based on the actual characteristic parameters.

[0008] The constraint scheduling module is used to input the liquid level prediction sequence, the inflow water flow prediction sequence, the current electricity price period information, and the water pump health score into the deep reinforcement learning scheduling model, and generate scheduling actions under the condition of meeting flood control safety constraints.

[0009] An adaptive correction module is used to adaptively correct the scheduling action based on the water pump health score and the actual characteristic parameters, so as to obtain the corrected scheduling action.

[0010] The simulation arbitration module is used to input the corrected scheduling action into a preset pipeline hydrodynamic simulation model for deduction, and obtain the predicted forebay liquid level after executing the corrected scheduling action; when the predicted forebay liquid level meets the preset safety conditions, the corrected scheduling action is output as the final control command; otherwise, the preset safety strategy is activated to generate the final control command.

[0011] The instruction execution module is used to send the final control instruction to the pump station controller to execute the water pump scheduling.

[0012] Preferably, in the data acquisition and joint prediction module, the step of acquiring multi-source operational data of the pumping station specifically includes:

[0013] Acquire gridded rainfall forecast data, daily time-of-use electricity price curves, forebay liquid level data, inflow and outflow data, operating status data of each water pump, and water pump electrical parameter data;

[0014] The acquired data is timestamped to generate time-series feature vectors with uniform resolution.

[0015] Preferably, in the data acquisition and joint prediction module, the step of generating a liquid level prediction sequence and an inflow rate prediction sequence for a future preset time period through a joint time series prediction model specifically includes:

[0016] The aligned temporal feature vector is input into the joint temporal prediction model;

[0017] The joint temporal prediction model employs a temporal convolutional network to extract long-distance temporal dependencies from the temporal feature vectors and outputs temporal feature representations.

[0018] Based on the temporal feature representation, the predicted liquid level sequence and the predicted inflow water sequence for the future preset time period are output.

[0019] Preferably, in the health identification module, the step of online identification of the actual characteristic parameters of each water pump specifically includes:

[0020] From the electrical parameter data of the water pump, the flow rate data and power data during steady-state operation of the water pump are selected as valid samples;

[0021] An online identification algorithm is used to fit a flow-power characteristic curve to the effective samples, and the characteristic curve is represented by a quadratic polynomial.

[0022] Based on the characteristic curve parameters obtained from the fitting, the actual characteristic parameters of each water pump are determined.

[0023] Preferably, in the health identification module, the step of determining the health score of each water pump based on the actual characteristic parameters specifically includes:

[0024] The actual efficiency of each pump is calculated based on the actual characteristic parameters, and the efficiency decay rate is determined.

[0025] Obtain the cumulative runtime of each water pump;

[0026] Based on the efficiency decay rate and the cumulative running time, the health score of each water pump is determined;

[0027] The health score is smoothed using an exponential moving average to eliminate transient fluctuations.

[0028] Preferably, in the constraint scheduling module, the step of generating scheduling actions under the condition of satisfying flood control safety constraints specifically includes:

[0029] Construct a state vector for inputting the deep reinforcement learning scheduling model. The state vector includes the liquid level prediction sequence, the inflow water prediction sequence, the current electricity price period label, the countdown to the next electricity price period, the current operating status of each water pump, and the health score of each water pump.

[0030] The state vector is input into the deep reinforcement learning scheduling model, which is trained with energy-saving reward as the optimization objective and flood control safety as the hard constraint.

[0031] The scheduling model outputs start / stop commands and operating frequency settings for each water pump based on the state vector.

[0032] Preferably, in the adaptive correction module, the step of obtaining the corrected scheduling action specifically includes:

[0033] When the scheduling action requires the water pump to be started and there are multiple water pumps available, a comprehensive selection score for each available water pump is calculated. The score is determined based on the actual efficiency of the water pump, its health score, and the cumulative running time of the day.

[0034] Select the water pump with the highest comprehensive score to start;

[0035] Replace the energy consumption assessment benchmark in the scheduling action with the actual characteristic parameters of the water pump from the rated characteristic parameters of the water pump.

[0036] Output the corrected scheduling action.

[0037] Preferably, in the simulation arbitration module, the step of obtaining the predicted forebay level after executing the corrected scheduling action specifically includes:

[0038] Construct a one-dimensional hydrodynamic simulation model covering the pump station forebay and its upstream local pipe section;

[0039] Using the current real-time forebay level and inflow / outflow rates as the initial state, the drainage volume corresponding to the corrected scheduling action as the boundary condition, and the current rainfall forecast as the input, the simulation model is driven to deduce the system state evolution trajectory within a future preset time period.

[0040] The predicted forebay level is extracted from the evolution trajectory.

[0041] Preferably, in the simulation arbitration module, when the predicted forebay level meets the preset safety conditions, the corrected scheduling action is output as the final control command; otherwise, the step of using the preset safety strategy to generate the final control command specifically includes:

[0042] When the predicted forebay level meets the preset safety conditions, the corrected scheduling action will be output as the final control command.

[0043] Otherwise, a preset security policy is activated to generate the final control command, which includes at least a regular veto policy and an emergency veto policy.

[0044] Preferably, the conventional veto strategy and the emergency veto strategy specifically include:

[0045] When the conventional veto strategy is triggered, the minimum drainage flow required to maintain the stability of the liquid level is calculated based on the current liquid level and the predicted inflow sequence, and the final control command is generated by back-calculating the pump operating parameters accordingly.

[0046] When the emergency veto strategy is triggered, a control command is generated to start all water pumps to run at full load.

[0047] The present invention has the following beneficial effects:

[0048] 1. This invention achieves multi-objective collaborative optimization of flood control safety and energy conservation at pumping stations through a collaborative working mechanism of joint time-series prediction, deep reinforcement learning-constrained scheduling, and pipeline hydrodynamic simulation arbitration. The joint time-series prediction model provides forward-looking liquid level and inflow predictions, the deep reinforcement learning scheduling model generates scheduling actions with flood control safety as a hard constraint, and the simulation arbitration model performs safety simulations and verifications of the scheduling actions. This effectively solves the problems of separation between prediction and scheduling and difficulty in balancing safety and energy conservation in existing technologies.

[0049] 2. In this invention, by identifying the actual characteristic parameters of the water pump online and generating a health score, the water pump with high efficiency, good health and reasonable running time is selected first to perform the drainage task in the scheduling correction stage. This realizes dynamic perception and active protection of the equipment health status, extends the service life of the water pump and reduces the operation and maintenance cost.

[0050] 3. In this invention, a lightweight pipeline hydrodynamic simulation model and a hierarchical preset safety strategy provide an independent physical verification and safety fallback mechanism for deep reinforcement learning scheduling actions. When the simulated liquid level does not meet the safety conditions, the AI ​​action is automatically rejected and the safety strategy is activated, significantly improving the system's reliability under extreme conditions. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the architecture of a deep learning-based smart water management system proposed in this invention. Detailed Implementation

[0052] The technical solutions in 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.

[0053] This invention provides a smart water management system based on deep learning, such as... Figure 1 As shown, it includes:

[0054] The data acquisition and joint prediction module is used to acquire multi-source operation data of the pumping station and generate liquid level prediction sequences and inflow flow prediction sequences for a preset time period through a joint time series prediction model.

[0055] Furthermore, in the data acquisition and joint prediction module, the specific steps for obtaining multi-source operational data of the pumping station include:

[0056] Acquire gridded rainfall forecast data, daily time-of-use electricity price curves, forebay liquid level data, inflow and outflow data, operating status data of each water pump, and water pump electrical parameter data;

[0057] The acquired data is timestamped to generate time-series feature vectors with uniform resolution.

[0058] Furthermore, in the data acquisition and joint prediction module, the steps of generating the liquid level prediction sequence and the inflow rate prediction sequence for a preset future time period through the joint time series prediction model specifically include:

[0059] The aligned temporal feature vectors are input into the joint temporal prediction model;

[0060] The joint temporal prediction model employs a temporal convolutional network to extract long-distance temporal dependencies from the temporal feature vectors and outputs temporal feature representations.

[0061] Based on temporal feature representation, the system outputs the predicted liquid level sequence and the predicted inflow water sequence for a future preset time period.

[0062] Specifically, gridded rainfall forecast data for the pumping station area is periodically acquired via an application programming interface (API) provided by the meteorological department. This rainfall forecast data includes the rainfall intensity, duration, and coverage for the next six to twelve hours. The time-of-use (TOU) electricity price curve for the day is obtained through a public interface provided by the power company or through manual input. This TOU price curve includes the start and end times of peak, normal, and valley periods, as well as the corresponding electricity prices. Real-time data on the forebay level, inflow and outflow rates, and the operating status of each pump are read from the pumping station's SCADA system via the OPC protocol. This operating status data includes the pump's start / stop status, current operating frequency, and cumulative operating time. Real-time current and power data for each pump are collected from the pump's frequency converter or the smart meter in the distribution cabinet at a preset sampling frequency of no less than once per minute, serving as the pump's electrical parameter data.

[0063] The acquired rainfall forecast data, time-of-use electricity price curves, forebay liquid level data, influent and effluent flow rate data, and pump electrical parameter data are timestamped and aligned. Due to differences in sampling frequency and transmission delay among the various data sources, linear interpolation or nearest-neighbor interpolation methods can be used to unify the data sequences to the same time base, generating a time-series feature vector with uniform resolution. A preferred time resolution is one sampling point per minute. During the interpolation process, the rainfall forecast data is processed using a time-decay weighting method, assigning higher interpolation weights to forecast data more recent than the current time.

[0064] The specific implementation process of generating the liquid level prediction sequence and the inflow rate prediction sequence for a future preset time period through a joint time series prediction model is as follows: The aligned time series feature vector is input into the joint time series prediction model. This time series feature vector... Recorded as:

[0065] ;

[0066] in, Indicates the current moment. Indicates the length of the backtracking time window. This represents the feature dimension at each time step.

[0067] The joint temporal prediction model employs a temporal convolutional network as its core architecture. The temporal convolutional network extracts long-range temporal dependencies from the temporal feature vectors through its internal structure and outputs a temporal feature representation, denoted as . Where h represents the hidden layer feature dimension.

[0068] Based on temporal feature representation, a fully connected output layer generates in parallel a predicted sequence of liquid level and inflow rate for a predetermined future time period. Specifically, the output layer uses two independent linear transformations to map to the predicted liquid level and flow rate values, respectively:

[0069] ;

[0070] ;

[0071] in, and Let represent the predicted liquid level and the predicted flow rate at the t-th time step in the future, respectively. , , , For trainable weight matrices and bias terms, This represents the feature vector corresponding to the time step in the time series feature representation. The preset future time period can be set to six hours, and the prediction step size is fifteen minutes.

[0072] The training process of the joint time series prediction model uses the mean squared error loss function, the optimizer is Adam, and the initial learning rate is set to [value missing]. The learning rate decay strategy was employed. The training data consisted of a supervised sample set constructed from the historical operating data of the pumping station. The input was a historical time-series feature vector, and the labels were the actual liquid level sequence and flow rate sequence for the corresponding future time period.

[0073] By following the steps above, we can obtain the predicted liquid level sequence and the predicted water flow rate sequence for the future preset time period, providing a forward-looking data foundation for subsequent scheduling decisions.

[0074] The health identification module is used to identify the actual characteristic parameters of each water pump online by using the water pump electrical parameter data from multiple sources of operating data, and to determine the health score of each water pump based on the actual characteristic parameters.

[0075] Furthermore, in the health assessment module, the steps for online identification of the actual characteristic parameters of each water pump specifically include:

[0076] From the electrical parameter data of the water pump, the flow rate data and power data during steady-state operation of the water pump are selected as valid samples;

[0077] An online identification algorithm is used to fit the flow-power characteristic curve to the effective samples. The characteristic curve is represented by a quadratic polynomial.

[0078] Based on the characteristic curve parameters obtained from the fitting, the actual characteristic parameters of each water pump are determined.

[0079] Furthermore, in the health identification module, the steps for determining the health score of each water pump based on actual characteristic parameters specifically include:

[0080] Calculate the actual efficiency of each pump based on its actual characteristic parameters, and determine the efficiency decay rate.

[0081] Obtain the cumulative runtime of each water pump;

[0082] The health score of each pump is determined based on the efficiency decay rate and cumulative running time.

[0083] The health score is smoothed using an exponential moving average to eliminate the interference of instantaneous fluctuations.

[0084] Specifically, from the electrical parameter data of the water pumps, flow rate and power data during steady-state operation are selected as valid samples. The criteria for steady-state operation are: the fluctuation range of the water pump operating frequency within a continuous preset time is less than a first preset threshold, and the fluctuation range of the inlet and outlet water flow is less than a second preset threshold. A preferred preset time is five minutes. An online identification algorithm is used to perform online fitting of the flow rate-power characteristic curve of each water pump based on the valid samples. The flow rate-power characteristic curve is represented by a quadratic polynomial:

[0085] ;

[0086] in, Indicates the operating power of the water pump. Indicates the water pump's outlet flow rate. , , These are the parameters of the characteristic curve to be identified.

[0087] The online identification algorithm employs recursive least squares with a forgetting factor. The recursive formula for recursive least squares is:

[0088] ;

[0089] ;

[0090] ;

[0091] in, This represents the parameter estimation vector at the k-th sampling time. ; This represents the measured power value at the k-th sampling time; This represents the regression vector at the k-th sampling time. ; Represents the gain matrix; Represent the covariance matrix; This represents the forgetting factor, with a value ranging from 0.95 to 0.99. Represents the identity matrix.

[0092] By updating the parameter estimation vector online using the recursive least squares method, the actual characteristic parameters of each water pump can be obtained. , , The introduction of a forgetting factor enables the identification algorithm to track the slow degradation of pump performance over time, with recently collected data being given higher weight.

[0093] The specific implementation process for determining the health score of each water pump based on actual characteristic parameters is as follows:

[0094] The actual efficiency of each pump is calculated based on its actual characteristic parameters. The formula for calculating the actual efficiency is:

[0095] ;

[0096] in, Indicates actual efficiency. Indicates fluid density, Represents gravitational acceleration. Indicates the pump head. This is the power value calculated based on the actual characteristic parameters.

[0097] The efficiency decay is determined based on the actual efficiency and the corresponding rated efficiency of the water pump. The formula for calculating the efficiency decay is:

[0098] ;

[0099] in, This represents the degree of efficiency decay, with a value ranging from zero to one. This indicates the rated efficiency of the water pump under the same operating conditions.

[0100] Obtain the cumulative runtime of each water pump, denoted as . The cumulative runtime can be obtained from the operating status data of each pump in the SCADA system. Based on the efficiency decay and cumulative runtime, the health score of each pump is determined. The formula for calculating the health score is:

[0101] ;

[0102] in, This represents the health score, with a value ranging from zero to one. The higher the value, the better the health status of the water pump. Indicates the degree of efficiency decay; Indicates the cumulative runtime; The weighting coefficient represents the degree of efficiency decay, and its value ranges from zero to one; This represents the attenuation coefficient indicating the cumulative runtime. A preferred value is... =0.6, = per hour.

[0103] The health score is smoothed using an exponential moving average to eliminate transient fluctuations. The smoothing formula is as follows:

[0104] ;

[0105] in, This represents the smoothed health score after the k-th calculation. This represents the original health score calculated in the k-th iteration. This represents the smoothing coefficient, with a value ranging from zero to one. A preferred value is... =0.1.

[0106] By following the steps above, the health score of each water pump can be obtained. This score comprehensively reflects the degree of degradation in the operating efficiency of the water pump and the degree of accumulation of mechanical wear, providing a quantitative basis for equipment health protection in subsequent scheduling decisions.

[0107] The constraint scheduling module is used to input the liquid level prediction sequence, the inflow water flow prediction sequence, the current electricity price period information, and the water pump health score into the deep reinforcement learning scheduling model, and generate scheduling actions under the condition of meeting the flood control safety constraints.

[0108] Furthermore, in the constraint scheduling module, the steps for generating scheduling actions under the condition of satisfying flood control safety constraints specifically include:

[0109] Construct the state vector of the input deep reinforcement learning scheduling model. The state vector includes the liquid level prediction sequence, the inflow water flow prediction sequence, the current electricity price period label, the countdown to the next electricity price period, the current operating status of each water pump, and the health score of each water pump.

[0110] The state vector is input into the deep reinforcement learning scheduling model, which is trained with energy-saving rewards as the optimization objective and flood control safety as the hard constraint.

[0111] The scheduling model is based on state vectors and outputs start / stop commands and operating frequency settings for each pump.

[0112] Specifically, the state vector of the input deep reinforcement learning scheduling model is constructed. The state vector includes the following elements: the liquid level prediction sequence and the inflow rate prediction sequence obtained from the data acquisition and joint prediction steps, denoted as:

[0113] ;

[0114] ;

[0115] in, This represents the predicted liquid level value at the i-th time step in the future. Let N represent the predicted inflow rate at the i-th time step in the future, and let N represent the number of prediction steps.

[0116] The current electricity price time period label is obtained from the data collection and joint forecasting steps, and is denoted as... Electricity price period labels use discrete encoding; for example, peak hours are coded as zero, normal hours as one, and off-peak hours as two. The countdown time from the current moment to the next electricity price period is calculated and denoted as... The unit is minutes. The current operating status of each pump is obtained from the data acquisition and joint forecasting steps, including start / stop status, current operating frequency, and cumulative daily operating time. For the i-th pump, its current operating status is denoted as:

[0117] ;

[0118] in, Indicates the start / stop status, with zero representing stop and one representing start; Indicates the current operating frequency; This indicates the total runtime for the day.

[0119] The health score of each water pump is obtained from the health identification step. For the i-th water pump, it is denoted as... The above elements are concatenated into a unified state vector, denoted as:

[0120] ;

[0121] Where M represents the total number of water pumps in the pumping station.

[0122] The state vector is input into the deep reinforcement learning scheduling model. The deep reinforcement learning scheduling model is trained using a reinforcement learning framework based on constrained Markov decision processes. During training, flood control safety is set as a hard constraint that must be met, and energy conservation and emission reduction are set as the optimization objective. The training optimization problem of the deep reinforcement learning scheduling model is expressed as:

[0123] ;

[0124] ;

[0125] in, Indicates the scheduling strategy. This represents the trajectory generated by the strategy. This represents the discount factor, which is preferably 0.99 in this scheme. Represents the energy-saving reward function. This represents the flood control safety cost function. This represents the safety constraint threshold, which is preferably zero in this scheme, indicating that the expected value of the flood control safety cost is zero. Indicates the scheduling strategy The expected value is obtained from the sampled trajectory.

[0126] The energy-saving reward function is defined as the electricity cost saved compared to the baseline full-load operation strategy:

[0127] ;

[0128] in, This represents the electricity cost of the benchmark strategy. This indicates the total energy consumption corresponding to the current scheduling action. This indicates the unit electricity price corresponding to the current electricity price period.

[0129] The flood control safety cost function is defined as the degree to which the liquid level in the forebay exceeds the safety threshold:

[0130] ;

[0131] in, This indicates the current level of the forebay. This indicates a preset safety threshold.

[0132] The training process employs the Lagrange relaxation method to transform the constrained optimization problem into an unconstrained minimax problem. Lagrange multipliers are introduced. ≥0, construct the Lagrange function:

[0133] ;

[0134] During training, the policy network parameters are updated by maximizing the Lagrange function, and the Lagrange multipliers are updated by gradient ascent, so that the policy gradually learns to maximize energy-saving benefits while satisfying the liquid level safety constraints.

[0135] The policy network of the deep reinforcement learning scheduling model adopts an Actor-Critic architecture. The Actor network outputs the probability distribution of scheduling actions, and the Critic network outputs state value estimates. Scheduling actions include start / stop commands for each pump and the inverter frequency setpoint for each running pump, denoted as:

[0136] ;

[0137] in, This represents the start / stop command for the i-th water pump, where zero indicates stop and one indicates start; ∈ This represents the frequency setting value of the inverter for the i-th water pump. and These are the minimum and maximum frequencies allowed by the frequency converter, respectively, with a preferred range of 30 Hz to 50 Hz.

[0138] The deep reinforcement learning scheduling model, based on the input state vector, outputs start / stop commands and operating frequency settings for each water pump under the condition of satisfying flood control safety constraints, thus generating scheduling actions.

[0139] The adaptive correction module is used to adaptively correct the scheduling actions based on the pump health score and actual characteristic parameters, and obtain the corrected scheduling actions.

[0140] Furthermore, in the adaptive correction module, the steps for obtaining the corrected scheduling action specifically include:

[0141] When a scheduling action requires the pump to be started and there are multiple pumps available, a comprehensive selection score for each pump is calculated. The score is determined based on the pump's actual efficiency, health score, and cumulative running time for the day.

[0142] Select the water pump with the highest comprehensive score to start;

[0143] Replace the rated characteristic parameters of the water pump with the actual characteristic parameters in the energy consumption assessment of the scheduling action;

[0144] Output the corrected scheduling action.

[0145] Specifically, the scheduling actions output by the constraint scheduling steps are parsed to obtain the start / stop commands for each water pump and the frequency setting value of the frequency converter. When a scheduling action requires starting a water pump, and there are multiple selectable water pumps in the pumping station, start / stop optimization correction is performed. Selectable water pumps refer to those currently stopped, without fault alarms, not locked by maintenance, and possessing normal start-up conditions. For each selectable water pump, its comprehensive selection score is calculated. The comprehensive selection score is determined based on the pump's actual efficiency, health score, and cumulative daily runtime. The formula for calculating the comprehensive selection score is:

[0146] ;

[0147] in, This represents the overall selection score for the i-th water pump, with a value ranging from zero to one. The actual efficiency of the i-th water pump is calculated by the health assessment step. This represents the health score of the i-th water pump after smoothing. This represents the cumulative daily runtime of the i-th water pump; This indicates the maximum cumulative daily operating time allowed for the water pump; a preferred value is eight hours. , , These are the actual efficiency weighting coefficient, the health score weighting coefficient, and the runtime weighting coefficient, respectively, satisfying... + + =1. A preferred weight configuration is: =0.4, =0.4, =0.2.

[0148] A higher overall selection score indicates that the pump is more suitable for drainage tasks under the current operating conditions. The pump with the highest overall selection score is selected from all available pumps for startup, and the start / stop commands for other pumps specified in the original scheduling action are cancelled or adjusted accordingly. After completing the start / stop optimization correction, a dynamic correction of the energy consumption baseline is performed. The rated characteristic parameters of the pumps used in the expected energy consumption calculation in the scheduling action are replaced online with the actual characteristic parameters obtained in the health identification step. Specifically, for the i-th pump, its power-flow characteristic curve is replaced from the rated curve to the actual characteristic curve:

[0149] ;

[0150] in, , , The actual characteristic parameters obtained by online identification of the i-th water pump are: This refers to the pump's outlet flow rate. Calculating the expected energy consumption of scheduling actions based on actual characteristic parameters can more accurately reflect the pump's true energy consumption level under its current health condition, ensuring the long-term effectiveness of the energy-saving performance evaluation benchmark.

[0151] The output is the corrected scheduling action after start-stop optimization correction and energy consumption benchmark dynamic correction. The corrected scheduling action includes the pump start-stop command selected by optimization and the frequency setting value of the inverter for each operating pump. Its data format is consistent with the scheduling action output by the constraint scheduling step.

[0152] Through the above steps, the adaptive matching between scheduling actions and equipment health status is achieved. While meeting the needs of flood control scheduling, priority is given to scheduling water pumps with good health status and high operating efficiency to perform drainage tasks, so that water pumps with poor health status or long operating time on the day can have a chance to rest and recuperate, thereby extending the overall service life of the water pumps.

[0153] The simulation arbitration module is used to input the corrected scheduling action into a preset pipeline hydrodynamic simulation model for deduction and obtain the predicted forebay liquid level after executing the corrected scheduling action. When the predicted forebay liquid level meets the preset safety conditions, the corrected scheduling action is output as the final control command; otherwise, the preset safety strategy is activated to generate the final control command.

[0154] Furthermore, in the simulation arbitration module, the specific steps for obtaining the predicted forebay level after executing the corrected scheduling action include:

[0155] Construct a one-dimensional hydrodynamic simulation model covering the pump station forebay and its upstream local pipe section;

[0156] Using the current real-time forebay level and inflow / outflow rates as the initial state, the drainage volume corresponding to the corrected scheduling action as the boundary condition, and the current rainfall forecast as the input, the simulation model is driven to deduce the system state evolution trajectory within a preset future time period.

[0157] Extract the predicted forebay level from the evolutionary trajectory.

[0158] Furthermore, in the simulation arbitration module, when the predicted forebay level meets the preset safety conditions, the corrected scheduling action is output as the final control command; otherwise, the steps for generating the final control command using the preset safety strategy specifically include:

[0159] When the predicted forebay level meets the preset safety conditions, the corrected scheduling action will be output as the final control command.

[0160] Otherwise, the default security policy is used to generate the final control command. The default security policy includes at least a regular veto policy and an emergency veto policy.

[0161] Furthermore, the regular veto strategy and the emergency veto strategy specifically include:

[0162] When the conventional veto strategy is triggered, the minimum drainage flow required to maintain the stability of the liquid level is calculated based on the current liquid level and the predicted inflow sequence, and the final control command is generated by back-calculating the pump operating parameters accordingly.

[0163] When the emergency veto strategy is triggered, a control command is generated to start all water pumps to run at full load.

[0164] Specifically, a one-dimensional hydrodynamic simulation model covering the pump station forebay and its upstream local pipe section is constructed. This model is based on engineering data such as pump station design drawings, pipeline geographic information system data, and downstream boundary conditions, with the number of spatial discrete units controlled within the range of ten to fifty. The governing equations of the one-dimensional hydrodynamic simulation model adopt the spreading wave equation, neglecting the inertial term. The spreading wave equation is simplified from the one-dimensional Saint-Venant equations, and its continuity equation and momentum equation are as follows:

[0165] ;

[0166] ;

[0167] in, Indicates the cross-sectional area of ​​the water passage. Indicates time, Indicates the flow rate of the pipe section. Represents the coordinates along the pipe segment direction. This represents the lateral inflow rate per unit length. Indicates water depth. Indicates the slope of the pipe bottom. This represents the friction drop. The friction drop is calculated using the Manning formula:

[0168] ;

[0169] Where n represents the pipe wall roughness coefficient and R represents the hydraulic radius.

[0170] The finite difference method is used to discretize and solve the diffusion wave equation. The time step is adaptively adjusted according to the Courant-Friedrichs-Lewy condition to ensure numerical stability. The initial state is based on the currently acquired real-time forebay liquid level and influent / outfluent flow rate data. The boundary conditions are the pump discharge volume corresponding to the corrected scheduling action, and the current rainfall forecast data is used as the external input to drive the one-dimensional hydrodynamic simulation model for extrapolation. The extrapolation duration is a preset future time period; a preferred extrapolation duration is fifteen minutes.

[0171] The simulation process yields the system state evolution trajectory over a predetermined future time period. This trajectory includes the water depth and flow rate values ​​of each discrete node at each time step. The water depth sequence of the corresponding node in the pump station forebay is extracted from the system state evolution trajectory and used as the predicted forebay level, denoted as:

[0172] ;

[0173] in, This represents the predicted forepool level at the k-th time step, where K represents the total number of simulation time steps.

[0174] After the aforementioned lightweighting process, the computation time for simulating a 15-minute operating condition is controlled to within 0.1 seconds, achieving over 150 times the real-time simulation capability. The one-dimensional hydrodynamic simulation model possesses online self-correction capabilities based on ensemble Kalman filtering. When the deviation between the simulated liquid level and the actual sensor reading exceeds a preset threshold, the model's key parameters are automatically corrected by assimilating real-time monitoring data. The state update formula for ensemble Kalman filtering is:

[0175] ;

[0176] in, This represents the predicted state vector of the i-th set member. Let K represent the analysis state vector of the i-th set member, and let K represent the Kalman gain matrix. Let H represent the observation vector, and let H represent the observation operator. This indicates an observed disturbance.

[0177] When the predicted forebay level meets the preset safety conditions, the corrected scheduling action will be output as the final control command; otherwise, the specific implementation process of generating the final control command using the preset safety strategy is as follows: Extract the highest value of the predicted forebay level during the simulation period from the system state evolution trajectory, and denot it as:

[0178] ;

[0179] Determine whether the predicted forebay level meets the preset safety conditions. The preset safety conditions are:

[0180] ;

[0181] in, This indicates a preset safety threshold, with a preferred value being 80% of the highest designed water level in the pump station forebay. This represents the safety factor, which ranges from 0.9 to 1.0 and is used to reserve a safety margin under model uncertainty conditions.

[0182] When the predicted forebay level meets the preset safety conditions, the arbitration decision is adopted, and the revised scheduling action is output as the final control command. When the predicted forebay level does not meet the preset safety conditions, the arbitration decision is rejected, and the preset safety strategy is activated to generate the final control command. The preset safety strategy includes at least a regular rejection strategy and an emergency rejection strategy.

[0183] The trigger condition for the conventional veto strategy is: the predicted forebay level exceeds a preset safety threshold, and the current real-time forebay level is below the overflow threshold. When the conventional veto strategy is triggered, the minimum drainage flow rate required to maintain level stability is calculated based on the current real-time forebay level and the predicted inflow rate sequence. The formula for calculating the minimum drainage flow rate is:

[0184] ;

[0185] in, Indicates the minimum drainage flow rate. This represents the predicted average inflow rate over a predetermined time period. Indicates the target storage capacity. Indicates the current storage capacity. Indicates the duration of the control cycle.

[0186] Based on the minimum drainage flow rate and the actual characteristic parameters of each pump, the operating parameters of the pumps that meet the minimum drainage flow rate are deduced, and the final control command is generated. The emergency veto strategy is triggered when: the predicted forebay level exceeds the preset safety threshold, and the current real-time forebay level has reached or exceeded the overflow critical value; or the system communication is interrupted or the sensor data is abnormal. When the emergency veto strategy is triggered, a control command is generated that includes starting all available pumps and setting each pump to full-load operation, which serves as the final control command. When the arbitration result is a veto, the system feeds back the veto record to the experience pool of the deep reinforcement learning scheduling model. The veto record includes the reason for the veto and the predicted level curve, which is used by the deep reinforcement learning scheduling model for offline fine-tuning, enabling the deep reinforcement learning scheduling model to gradually learn to avoid action strategies that are easily vetoed. When the sensor data is abnormal or the communication is interrupted, the deduction and arbitration of the one-dimensional hydrodynamic simulation model are skipped, and the emergency veto strategy is executed directly.

[0187] The instruction execution module is used to send the final control instructions to the pump station controller to execute the water pump scheduling.

[0188] Specifically, the final control commands are obtained from the simulation arbitration process. These final control commands include start / stop commands for each water pump and the inverter frequency setpoint for each operating water pump. The data structure of the final control commands is as follows:

[0189] ;

[0190] in, This represents the final start / stop command for the i-th water pump, where zero indicates stop and one indicates start; This represents the final inverter frequency setting value for the i-th water pump, and M represents the total number of water pumps in the pumping station.

[0191] The final control commands are sent to the programmable logic controller (PLC) of the pumping station via the standard OPC protocol. The OPC protocol uses a client-server architecture, with the smart water management system acting as the OPC client and the PLC acting as the OPC server. Each data item in the final control command is mapped one-to-one with a data point in the PLC. Start / stop commands are mapped to the digital output points of the PLC to control the engagement and disengagement of the pump contactors; the inverter frequency setpoint is mapped to the analog output points of the PLC to control the inverter's frequency command. Upon receiving the final control command, the PLC executes the following control actions:

[0192] For each water pump, if the final start / stop command is one and the current state is stopped, the programmable logic controller (PLC) outputs a closed signal to the corresponding contactor to start the pump; if the final start / stop command is zero and the current state is running, the PLC outputs an open signal to the corresponding contactor to stop the pump; if the final start / stop command matches the current state, the original state remains unchanged. For pumps in operation, the PLC converts the final inverter frequency setpoint into a corresponding analog signal and outputs it to the inverter's analog input terminal, which the inverter uses to adjust the pump motor's operating frequency. After executing the final control command, the PLC returns an execution confirmation signal to the smart water management system via the OPC protocol. Simultaneously, the SCADA system collects the execution results in real time, including the forebay level data, inlet and outlet flow rate data, and the actual operating status data of each pump. These execution results serve as the data input for the next control cycle, feeding back to the data acquisition and joint prediction steps, forming a complete closed loop of perception, prediction, decision-making, correction, arbitration, execution, and feedback.

[0193] The system is deployed in a three-phase go-live mode. The first phase is the recommended mode, where the system outputs final control commands but does not directly control the pumping station equipment; these commands are manually executed after operator review and confirmation. The second phase is the observation mode, where the system directly controls the pumping station equipment, but the operator can switch back to manual mode at any time using a one-click switching function. The third phase is the direct control mode, where the system switches to fully automatic operation mode after the recommended accuracy rate has continuously met the preset number of days. The final control commands output by the system are directly issued and executed without manual confirmation. A preferred preset number of days is seven days. When system communication is interrupted or the programmable logic controller (PLC) detects abnormal final control command data, the PLC automatically switches to local safety mode and executes preset local safety logic. Local safety logic includes: maintaining the current pump operating status, or automatically starting all pumps to full-load operation when the liquid level exceeds a preset limit threshold.

[0194] Through the above steps, the safe and reliable issuance and execution of the final control command are achieved, completing the intelligent closed-loop control of the entire process of pump station water pump scheduling.

[0195] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based intelligent water management system, characterized by, include: The data acquisition and joint prediction module is used to acquire multi-source operation data of the pumping station and generate liquid level prediction sequence and inflow rate prediction sequence for a future preset time period through a joint time series prediction model. The health identification module is used to identify the actual characteristic parameters of each water pump online through the water pump electrical parameter data of the multi-source operation data, and determine the health score of each water pump based on the actual characteristic parameters. The constraint scheduling module is used to input the liquid level prediction sequence, the inflow water flow prediction sequence, the current electricity price period information, and the water pump health score into the deep reinforcement learning scheduling model, and generate scheduling actions under the condition of meeting flood control safety constraints. An adaptive correction module is used to adaptively correct the scheduling action based on the water pump health score and the actual characteristic parameters, so as to obtain the corrected scheduling action. The simulation arbitration module is used to input the corrected scheduling action into a preset pipeline hydrodynamic simulation model for deduction, and obtain the predicted forebay liquid level after executing the corrected scheduling action; when the predicted forebay liquid level meets the preset safety conditions, the corrected scheduling action is output as the final control command; otherwise, the preset safety strategy is activated to generate the final control command. The instruction execution module is used to send the final control instruction to the pump station controller to execute the water pump scheduling.

2. The intelligent water management system based on deep learning according to claim 1, characterized in that, In the data acquisition and joint prediction module, the steps for acquiring multi-source operational data of the pumping station specifically include: Acquire gridded rainfall forecast data, daily time-of-use electricity price curves, forebay liquid level data, inflow and outflow data, operating status data of each water pump, and water pump electrical parameter data; The acquired data is timestamped to generate time-series feature vectors with uniform resolution. 3.The smart water management system based on deep learning according to claim 2, wherein, In the data acquisition and joint prediction module, the steps of generating the liquid level prediction sequence and the inflow rate prediction sequence for a preset future time period through the joint time series prediction model specifically include: The aligned temporal feature vector is input into the joint temporal prediction model; The joint temporal prediction model employs a temporal convolutional network to extract long-distance temporal dependencies from the temporal feature vectors and outputs temporal feature representations. Based on the temporal feature representation, the predicted liquid level sequence and the predicted inflow water sequence for the future preset time period are output. 4.The smart water management system based on deep learning according to claim 2, wherein, In the health assessment module, the steps for online identification of the actual characteristic parameters of each water pump specifically include: From the electrical parameter data of the water pump, the flow rate data and power data during steady-state operation of the water pump are selected as valid samples; An online identification algorithm is used to fit a flow-power characteristic curve to the effective samples, and the characteristic curve is represented by a quadratic polynomial. Based on the characteristic curve parameters obtained from the fitting, the actual characteristic parameters of each water pump are determined.

5. The intelligent water management system based on deep learning according to claim 4, characterized in that, In the health identification module, the step of determining the health score of each water pump based on the actual characteristic parameters specifically includes: The actual efficiency of each pump is calculated based on the actual characteristic parameters, and the efficiency decay rate is determined. Obtain the cumulative runtime of each water pump; Based on the efficiency decay rate and the cumulative running time, the health score of each water pump is determined; The health score is smoothed using an exponential moving average to eliminate transient fluctuations.

6. The intelligent water management system based on deep learning according to claim 1, characterized in that, In the constraint scheduling module, the steps for generating scheduling actions under the condition of satisfying flood control safety constraints specifically include: Construct a state vector for inputting the deep reinforcement learning scheduling model. The state vector includes the liquid level prediction sequence, the inflow water prediction sequence, the current electricity price period label, the countdown to the next electricity price period, the current operating status of each water pump, and the health score of each water pump. The state vector is input into the deep reinforcement learning scheduling model, which is trained with energy-saving reward as the optimization objective and flood control safety as the hard constraint. The scheduling model outputs start / stop commands and operating frequency settings for each water pump based on the state vector.

7. A deep learning-based smart water management system according to claim 1, characterized in that, In the adaptive correction module, the steps for obtaining the corrected scheduling action specifically include: When the scheduling action requires the water pump to be started and there are multiple water pumps available, a comprehensive selection score for each available water pump is calculated. The score is determined based on the actual efficiency of the water pump, its health score, and the cumulative running time of the day. Select the water pump with the highest comprehensive score to start; Replace the energy consumption assessment benchmark in the scheduling action with the actual characteristic parameters of the water pump from the rated characteristic parameters of the water pump. Output the corrected scheduling action.

8. A smart water management system based on deep learning according to claim 1, characterized in that, In the simulation arbitration module, the step of obtaining the predicted forebay level after executing the corrected scheduling action specifically includes: Construct a one-dimensional hydrodynamic simulation model covering the pump station forebay and its upstream local pipe section; Using the current real-time forebay level and inflow / outflow rates as the initial state, the drainage volume corresponding to the corrected scheduling action as the boundary condition, and the current rainfall forecast as the input, the simulation model is driven to deduce the system state evolution trajectory within a future preset time period. The predicted forebay level is extracted from the evolution trajectory.

9. A smart water management system based on deep learning according to claim 8, characterized in that, In the simulation arbitration module, when the predicted forebay level meets the preset safety conditions, the corrected scheduling action is output as the final control command; otherwise, the step of using the preset safety strategy to generate the final control command specifically includes: When the predicted forebay level meets the preset safety conditions, the corrected scheduling action will be output as the final control command. Otherwise, a preset security policy is activated to generate the final control command, which includes at least a regular veto policy and an emergency veto policy.

10. A deep learning-based smart water management system according to claim 9, characterized in that, The conventional veto strategy and the emergency veto strategy specifically include: When the conventional veto strategy is triggered, the minimum drainage flow required to maintain the stability of the liquid level is calculated based on the current liquid level and the predicted inflow sequence, and the final control command is generated by back-calculating the pump operating parameters accordingly. When the emergency veto strategy is triggered, a control command is generated to start all water pumps to run at full load.

Images