A deep learning-based drainage pump station scheduling method

Abstract

The application discloses a kind of based on deep learning's drainage pump station scheduling method, to solve the problem that the existing pump station adopts single station liquid level threshold start-stop or experience rule scheduling leads to the loss of upstream and downstream coupling of multiple pump stations, rainfall and receiving water body jacking is difficult to comprehensively predict, inspection well spills or overflow frequency is high, sewage treatment facilities inflow water quantity and water quality fluctuation is big and energy consumption is high, the application is by collecting pump station operation monitoring data, external influence data and drainage pipe network topology and pre-processing, calculate inflow and determine pump station classification and control priority, predict inflow, inflow pool liquid level, water quality and overflow risk using deep learning model, combined with downstream transport capacity and target pump station effluent flow fluctuation constraint to solve water pump start-stop threshold and frequency setting value and issue instruction, performance evaluation is carried out based on execution feedback and the model is retrained by updating sample, realize the technical effect of water level controllable, overflow reduction, inflow stable and energy consumption controllable.

Description

Technical Field

[0001] This invention relates to the field of automated control of urban drainage and sewage transportation, and in particular to a deep learning-based method for scheduling drainage pumping stations. Background Technology

[0002] Urban drainage systems typically consist of sewage pipe networks, stormwater pipe networks or combined sewer networks, manholes, and drainage pumping stations. As a crucial node in the pipe network's transport and lifting operations, the operation and scheduling of drainage pumping stations directly impact the safety of pipe network water levels, overflow control, and the stability of water quantity and quality entering sewage treatment facilities. Currently, with the development of online monitoring equipment, automated control systems, and scheduling platforms, drainage pumping stations are generally equipped with level gauges, flow meters, and pump operation status acquisition devices, and remote start / stop and frequency conversion control are achieved through programmable logic controllers (PLCs) or remote monitoring systems. In some areas, centralized scheduling systems have also been built to manage multiple pumping stations uniformly based on empirical rules or preset operating condition strategies. Simultaneously, some engineering practices are attempting to introduce pipe network hydraulic models or data-driven methods to assist in the prediction of short-term water inflows, supporting scheduling decisions.

[0003] The existing technology still has the following shortcomings:

[0004] 1. Existing scheduling mainly relies on single-station liquid level threshold start / stop or static empirical rules, lacking upstream and downstream coupling analysis based on drainage network topology and multi-pump station collaborative control mechanism, making it difficult to achieve global optimization in scenarios where multiple pump stations merge in parallel or upstream and downstream affect each other.

[0005] 2. The existing dispatching system does not adequately consider external influencing factors such as rainfall processes, inflow and infiltration of external water, and backwater levels in receiving water bodies. It usually lacks comprehensive prediction of short-term inflow, inlet pool level, and overflow risk, which leads to high water levels in the pipeline network, overflow of inspection wells, or high frequency of overflows during rainy days or special operating conditions.

[0006] 3. The existing scheduling system lacks overall coordination in controlling the influent stability and energy consumption of sewage treatment facilities, making it difficult to constrain fluctuations in pump station effluent while also taking energy conservation into account. Furthermore, the strategy relies on manual adjustments and lacks adaptive update capabilities based on execution feedback, making it difficult to achieve dynamic optimization operation.

[0007] Therefore, a drainage pumping station scheduling method that can overcome the shortcomings of the existing technology is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose a deep learning-based drainage pumping station scheduling method. This addresses the problems of existing technologies that generally rely on single-station liquid level threshold start / stop or empirical rule-based scheduling, lacking comprehensive prediction and coordinated control of upstream and downstream coupling relationships among multiple pumping stations, rainfall impacts, backwater effects from receiving water bodies, and changes in inflow volume and quality. This leads to problems such as high water level operation in the pipe network, frequent overflows or spills from manholes, large fluctuations in influent to sewage treatment facilities, and high energy consumption. The proposed method involves collecting and preprocessing pumping station operation monitoring data, external impact data, and drainage pipe network topology within a preset scheduling cycle; calculating the inflow rate based on the relationship between liquid level and effective volume; determining pumping station classification and control priorities; using a deep learning model to output predicted inflow rate, influent tank liquid level, water quality, and overflow risk prediction results in the predicted time domain; and solving for pump start / stop thresholds and frequency converter setpoints under liquid level constraints, downstream transport capacity constraints, and control priority constraints to form scheduling instructions. After execution, the method updates the monitoring samples based on feedback performance evaluation and retrains the model. This invention features controllable water level, reduced overflow, stable influent to wastewater treatment facilities, and controllable energy consumption, and can adaptively optimize based on operational feedback.

[0009] This invention provides a deep learning-based method for scheduling drainage pumping stations, comprising:

[0010] S1. Within a preset scheduling cycle, collect operational monitoring data, external impact data, and the topological relationship of the drainage network for each drainage pumping station. Timestamps are added, and preprocessing is performed to obtain processed monitoring data. S2. Based on the network topology, determine the upstream and downstream relationships and parallel relationships of each drainage pumping station. Using the influent level data and outfluent flow data from the processed monitoring data, and combining the correspondence between the influent level and effective volume, calculate the inflow rate for each drainage pumping station. S3. Classify the drainage pumping stations according to the upstream and downstream relationships to obtain a classification result. Identify the operating conditions based on the processed monitoring data to obtain operating condition identification results, determine the control priority of the drainage pumping stations, and identify the target drainage pumping stations connected to the sewage treatment facilities based on the network topology. S4. Extract temporal features within a preset historical time window from the processed monitoring data. Input these features, along with the operating condition identification results and the drainage pumping station control priorities, into a deep learning model to output the predicted time domain. S5. Based on the prediction results, the control priority of the drainage pumping station, and the current inlet tank level data, current outlet main flow data, and current pump operating status data from the post-treatment monitoring data, establish scheduling constraints to reduce overflow risk, reduce energy consumption, and ensure that the fluctuation range of the outlet flow of the target drainage pumping station does not exceed the preset threshold as the scheduling objectives. Solve for the pump start-stop control parameters and variable frequency operation setpoints of each drainage pumping station, and combine the pump start-stop control parameters and variable frequency operation setpoints into scheduling control commands; S6. Send the scheduling control commands to the automatic control system of the corresponding drainage pumping station to control the pump operation, and collect execution feedback monitoring data after the command is executed; S7. Evaluate the performance based on the execution feedback monitoring data to obtain the performance evaluation results. When the results do not meet the preset objectives, update the supervision samples based on the execution feedback monitoring data, and use the updated supervision samples to retrain the deep learning model to update the model parameters, so that the updated model parameters can be used in the next scheduling cycle.

[0011] Optionally, S1 includes:

[0012] Within the preset scheduling cycle, operation monitoring data, external impact data, and topological relationships of drainage pipe networks are collected for each drainage pumping station. The operation monitoring data includes inlet pool level data, outlet main pipe flow data, pump operation status data, and water quality data. The external impact data includes rainfall data and receiving water body level data.

[0013] The collected operational monitoring data and external impact data are each appended with a timestamp, and the timestamps are converted to a unified time base.

[0014] Based on a preset sampling period, time alignment is performed on the operation monitoring data and the external influence data after the additional timestamps are applied to obtain aligned time-series data;

[0015] The aligned time-series data are sequentially processed by missing value handling, outlier removal, and numerical normalization to obtain the processed monitoring data.

[0016] Optionally, S2 includes:

[0017] Based on the pipeline network topology, the upstream and downstream relationships and parallel relationships of each drainage pumping station are determined. The upstream and downstream relationships include the connectivity between each drainage pumping station along the flow direction of the drainage pipeline network, and the parallel relationships include the connectivity between at least two drainage pumping stations that converge into the same drainage pumping station or the same drainage pipeline.

[0018] Based on the inlet pool level data in the processed monitoring data, the inlet pool level at each time is converted into the corresponding effective volume using the correspondence between the inlet pool level and the effective volume, and the effective volume change rate is calculated based on the effective volume difference between adjacent time points and the time difference between adjacent time points.

[0019] The outflow rate at each time point is determined based on the outflow rate data of the main outflow pipe in the processed monitoring data;

[0020] The inflow rate of each drainage pumping station is calculated based on the outflow rate and the effective volume change rate, wherein the inflow rate is equal to the sum of the outflow rate and the effective volume change rate.

[0021] Optionally, S3 includes:

[0022] Based on the upstream and downstream relationships, the connection path between each drainage pumping station along the drainage network is determined, and based on the network topology, the drainage pumping station connected to the sewage treatment facility is determined as the target drainage pumping station.

[0023] The target drainage pumping station is identified as a first-level drainage pumping station, and the drainage pumping stations that have the upstream and downstream relationship with the first-level drainage pumping station and flow into the first-level drainage pumping station along the connection path are identified as second-level drainage pumping stations. The process continues to extend upstream along the connection path to obtain the drainage pumping station classification results.

[0024] Based on the rainfall data in the processed monitoring data, it is determined whether a rainfall process has occurred, and based on the water level data of the receiving water body and the liquid level data of the inlet pool in the processed monitoring data, it is determined whether there is an impact from the backwater of the receiving water body, thus obtaining the operating condition identification result.

[0025] Based on the classification results of the drainage pumping stations and combined with the inlet pool level data, outlet main flow data, pump operation status data and water quality data in the processed monitoring data, a corresponding control priority score is generated for each drainage pumping station, and the control priority of the drainage pumping station is determined according to the control priority score. The control priority score is at least related to the drainage pumping station level, the safety margin of the inlet pool level from the preset alarm level, the degree to which the outlet main flow is close to the downstream transport capacity, and the nature of the incoming water indicated by the water quality data.

[0026] And among the parallel drainage pumping stations corresponding to the parallel relationship, the scheduling order of the parallel drainage pumping stations is determined according to the control priority of the drainage pumping stations.

[0027] Optionally, S4 includes:

[0028] The processed monitoring data is used to extract sequence samples within a preset historical time window for each drainage pumping station. The sequence samples include, at each sampling time, the inlet pool level data, the outlet main flow rate data, the pump operating status data, the water quality data, the rainfall data, and the receiving water level data for that drainage pumping station.

[0029] The sequence samples, the operating condition identification results, and the drainage pump station control priority are jointly encoded into the input sequence of the deep learning model, and the deep learning model outputs the inflow prediction value, the inlet pool liquid level prediction value, and the water quality prediction value within the preset prediction time domain.

[0030] An overflow risk prediction value is generated based on the comparison between the predicted water level in the inlet tank and the preset alarm level within the preset prediction time domain. When the predicted water level in the inlet tank reaches or exceeds the preset alarm level, the overflow risk prediction value at the corresponding time is increased.

[0031] Optionally, S5 includes:

[0032] Based on the predicted values ​​of the inlet pool level and overflow risk in the prediction results, the level risk of each drainage pumping station within the preset prediction time domain is determined.

[0033] Based on the current inlet water level data, current outlet main flow data, and current pump operating status data in the processed monitoring data, the current operating conditions of each drainage pumping station are determined.

[0034] Based on the pipeline topology, the downstream transport capacity of each drainage pumping station is determined. This downstream transport capacity, along with the liquid level risk level and the control priority of the drainage pumping stations, is used to construct the scheduling constraints. These constraints include: the liquid level at the corresponding node of each drainage pumping station does not exceed the preset alarm liquid level; the total outflow of each drainage pumping station does not exceed its corresponding downstream transport capacity; and a control sequence that satisfies the control priority of the drainage pumping stations. Under the premise of satisfying the scheduling constraints, candidate control schemes are generated for each drainage pumping station. Each candidate control scheme includes the pump start-stop control parameters and variable frequency drive (VFD) operation setpoints for that drainage pumping station. The pump start-stop control parameters include a pump start liquid level threshold and a pump stop liquid level threshold. For each candidate... The control scheme calculates a scheduling target evaluation value, which is at least related to the overflow risk prediction value, energy consumption, and the fluctuation range of the outflow of the target drainage pumping station. The fluctuation range of the outflow of the target drainage pumping station is determined by the difference between the maximum and minimum values ​​of the outflow of the target drainage pumping station's main outlet pipe corresponding to the candidate control scheme within the preset prediction time domain, and the fluctuation range of the outflow of the target drainage pumping station is required not to exceed the preset threshold. From the candidate control schemes that meet the scheduling constraints and satisfy the requirement that the fluctuation range of the outflow of the target drainage pumping station does not exceed the preset threshold, the candidate control scheme with the optimal scheduling target evaluation value is selected, and the pump start-stop control parameters and variable frequency operation setpoints corresponding to the candidate control scheme are determined as the scheduling control command.

[0035] Optionally, S6 includes:

[0036] The scheduling control command is sent to the automatic control system of the corresponding drainage pumping station one by one. The automatic control system controls the start and stop of the pumps according to the pump start and stop control parameters in the scheduling control command, and controls the pumps to operate at variable frequency according to the variable frequency operation setpoint in the scheduling control command. After the scheduling control command is sent, the issuance time of the scheduling control command and the change time of the pump operation status data are recorded, and the effective time of the scheduling control command is determined by the issuance time and the change time. From the effective time, the actual inlet pool liquid level data, the actual outlet main flow data, the actual pump operation status data, and the actual water quality data are continuously collected within a preset feedback sampling period. Timestamps are added to the actual inlet pool liquid level data, the actual outlet main flow data, the actual pump operation status data, and the actual water quality data to obtain the execution feedback monitoring data.

[0037] Optionally, the S7 includes:

[0038] An overflow event is determined based on the comparison between the actual inlet water level data and the preset alarm level in the execution feedback monitoring data. Specifically, at the corresponding node of the same drainage pump station, the overflow event is defined as the start time when the actual inlet water level data changes from less than the preset alarm level to reaching or exceeding the preset alarm level, and the overflow event is defined as the end time when the actual inlet water level data changes from reaching or exceeding the preset alarm level to less than the preset alarm level.

[0039] The number of overflows is counted based on the start and end times of each overflow event, and the duration of each overflow event is calculated and accumulated to obtain the overflow duration.

[0040] Within the preset feedback sampling period, the safety margin of the critical node liquid level is calculated based on the difference between the maximum value of the actual inlet pool liquid level data at the corresponding node of each drainage pumping station in the execution feedback monitoring data and the preset alarm liquid level.

[0041] Within the preset feedback sampling period, the fluctuation range of the outflow of the target drainage pumping station is calculated based on the difference between the maximum and minimum values ​​of the actual outflow of the target drainage pumping station in the execution feedback monitoring data.

[0042] Within the preset feedback sampling period, the running time of each pump is determined based on the actual pump operating status data in the execution feedback monitoring data, and the energy consumption index is calculated based on the pump rated power and the running time.

[0043] The performance evaluation results are obtained by comparing the number of overflows, the duration of overflows, the safety margin of the liquid level at key nodes, the fluctuation range of the outflow of the target drainage pumping station, and the energy consumption indicators with the preset targets.

[0044] When the performance evaluation results do not meet the preset targets, the execution feedback monitoring data is preprocessed to update the processed monitoring data, and the inflow calculation value is calculated based on the inlet pool level data and the outlet main pipe flow data in the processed monitoring data.

[0045] The monitoring samples are updated based on the updated post-processing monitoring data, inflow calculation value, actual influent pool level data, and actual water quality data. The updated monitoring samples are then used to retrain the deep learning model to update the model parameters, so that the updated model parameters can be used in the deep learning prediction step of the next scheduling cycle.

[0046] Optionally, the deep learning model is a spatiotemporal graph neural network model. Based on the pipeline network topology, an adjacency matrix or graph edge set of pump station nodes is constructed. The sequence samples of adjacent pump station nodes are aggregated by graph convolution features and then time series modeling is performed to output the predicted inflow rate, the predicted level of the inlet pool, and the predicted water quality within the preset prediction time domain.

[0047] Optionally, calculating the total outflow of the target drainage pumping station for each of the candidate control schemes includes: establishing a hydraulic calculation model of the pipeline network based on the pipeline network topology, and inputting the pump operating status and frequency conversion operation setting value corresponding to the candidate control scheme as boundary conditions into the hydraulic calculation model of the pipeline network to obtain the sequence of total outflow of the target drainage pumping station in the preset prediction time domain.

[0048] The beneficial effects of this invention are:

[0049] 1. Achieve multi-pump station coordination and proactive risk control: Based on the drainage network topology, determine the upstream and downstream relationships and classification results of pump stations, and combine rainfall data, receiving water level data, and inflow and inlet pool liquid level predicted by deep learning to generate overflow risk prediction values. This enables coordinated scheduling before risks occur, reducing the probability of high water level operation in the network and reducing the number and duration of overflows or spills from manholes.

[0050] 2. Improve the stability of influent to sewage treatment facilities: Taking the target drainage pumping station connected to the sewage treatment facility as the control object, the constraint that the fluctuation of the outflow of the target drainage pumping station does not exceed the preset threshold is introduced in the scheduling solution. This can suppress large fluctuations in outflow, reduce the impact of influent water quantity and quality on sewage treatment facilities, and improve operational stability and treatment efficiency.

[0051] 3. Reduce energy consumption and have adaptive optimization capabilities: Under the premise of ensuring that the liquid level does not exceed the alarm, the water output does not exceed the downstream delivery capacity and the control priority order, the pump start-stop threshold and frequency conversion set value are optimized. Energy consumption is one of the objectives for optimization. At the same time, the supervision sample is updated based on the performance evaluation of the execution feedback and the deep learning model is retrained, so that the scheduling strategy can be continuously iterated and optimized according to the changes in the working conditions, reducing the reliance on human experience. Attached Figure Description

[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0053] Figure 1 This is a flowchart of a deep learning-based drainage pumping station scheduling method proposed in this invention. Detailed Implementation

[0054] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0055] refer to Figure 1 A deep learning-based method for scheduling drainage pumping stations includes:

[0056] S1. Within a preset scheduling cycle, collect operational monitoring data, external impact data, and the topological relationship of the drainage network for each drainage pumping station. Timestamps are added, and preprocessing is performed to obtain processed monitoring data. S2. Based on the network topology, determine the upstream and downstream relationships and parallel relationships of each drainage pumping station. Using the influent level data and outfluent flow data from the processed monitoring data, and combining the correspondence between the influent level and effective volume, calculate the inflow rate for each drainage pumping station. S3. Classify the drainage pumping stations according to the upstream and downstream relationships to obtain a classification result. Identify the operating conditions based on the processed monitoring data to obtain operating condition identification results, determine the control priority of the drainage pumping stations, and identify the target drainage pumping stations connected to the sewage treatment facilities based on the network topology. S4. Extract temporal features within a preset historical time window from the processed monitoring data. Input these features, along with the operating condition identification results and the drainage pumping station control priorities, into a deep learning model to output the predicted time domain. S5. Based on the prediction results, the control priority of the drainage pumping station, and the current inlet tank level data, current outlet main flow data, and current pump operating status data from the post-treatment monitoring data, establish scheduling constraints to reduce overflow risk, reduce energy consumption, and ensure that the fluctuation range of the outlet flow of the target drainage pumping station does not exceed the preset threshold as the scheduling objectives. Solve for the pump start-stop control parameters and variable frequency operation setpoints of each drainage pumping station, and combine the pump start-stop control parameters and variable frequency operation setpoints into scheduling control commands; S6. Send the scheduling control commands to the automatic control system of the corresponding drainage pumping station to control the pump operation, and collect execution feedback monitoring data after the command is executed; S7. Evaluate the performance based on the execution feedback monitoring data to obtain the performance evaluation results. When the results do not meet the preset objectives, update the supervision samples based on the execution feedback monitoring data, and use the updated supervision samples to retrain the deep learning model to update the model parameters, so that the updated model parameters can be used in the next scheduling cycle.

[0057] In this specific embodiment, S1 includes:

[0058] The preset scheduling period is set to The preset sampling period is set to During each scheduling cycle, operational monitoring data is collected for each drainage pumping station, and external impact data and the topological relationship of the drainage network are collected simultaneously. The operational monitoring data includes inlet pool level data, outlet main pipe flow data, pump operation status data, and water quality data. The external impact data includes rainfall data and receiving water body level data. The topological relationship of the drainage network is stored in a node-connection directed graph data structure and called according to version number during the scheduling cycle.

[0059] During data collection, each piece of operational monitoring data and external impact data is appended with a local timestamp from the data source. On the dispatching platform, the timestamp is converted to a unified timestamp using Beijing time (UTC+8) as a unified time reference. During the conversion, the clock deviation of each data source is obtained through the time synchronization service between the dispatching platform and the pump station automatic control system and is corrected before being written, thereby ensuring that different pump stations and external data sources can be compared on the same time axis.

[0060] After completing the unified timestamp, based on the sampling period Constructing aligned time series And map the original sampled values ​​from each data source to each The mapping rules are set as follows: For inlet tank level data, outlet main pipe flow rate data, and water quality data... The arithmetic mean within the time window is used as the alignment value for that sampling moment, for the water pump operating status data. The state that appears most frequently within the time window is taken as the alignment value for that sampling moment. If there is no valid original data in the corresponding time window, the sampling moment is recorded as missing.

[0061] Missing value handling is set as follows: when the length of consecutive missing values ​​for the same variable does not exceed [a certain value]... Linear interpolation is used to fill in the missing data when there are only a few sampling points. When the length of consecutive missing data exceeds a certain threshold, the missing data is filled in by the missing data. When sampling points are used, the most recent valid alignment value is retained until sampling is resumed;

[0062] Outlier removal is set to be executed by a joint rule based on engineering physical constraints and mutation constraints. Specifically: data on the inlet pool level exceeding the upper or lower limit of the pump station's inlet pool range is judged as outlier and replaced with the most recent valid alignment value; data on the outlet main flow exceeding the maximum designed conveying capacity of the pump station's outlet main is judged as outlier and replaced with the most recent valid alignment value; data on water quality exceeding the range of the monitoring instrument is judged as outlier and replaced with the most recent valid alignment value; when the pump operating status data shows a state combination that contradicts the pump station's electrical protection signal, the state at that sampling time is set to the pump stop state.

[0063] After handling missing values ​​and removing outliers, numerical normalization was performed on each variable for each pump station to generate processed monitoring data that could be directly input into subsequent deep learning models. The normalization employed min-max normalization and used a fixed parameter table. This fixed parameter table was generated by statistically analyzing the minimum and maximum values ​​of the corresponding variable for that pump station over the most recent 30 consecutive days of aligned data after missing and outlier handling, and was updated daily. Refresh once, normalize according to the formula:

[0064] ;

[0065] in, Indicates the sampling time The normalized value, Indicates the sampling time The original values ​​after alignment. Indicates by sampling period The generated first Each aligned sampling time, This represents the minimum value of the variable obtained from the aligned data of the most recent 30 consecutive days. This represents the maximum value of the variable obtained from the aligned data of the most recent 30 consecutive days, and when When To avoid division by zero and maintain numerical stability, the normalized inlet water level data, outlet main pipe flow data, water pump operating status data, water quality data, rainfall data, and receiving water body water level data are merged into processed monitoring data according to a unified timestamp order and written into the time series database of the scheduling platform.

[0066] In this specific embodiment, S2 includes:

[0067] The process is executed synchronously within each scheduling cycle based on the processed monitoring data and the topological relationship of the drainage network.

[0068] The topology of the drainage network is stored in the form of a directed connected list. Each record in the connected list includes an upstream node identifier, a downstream node identifier, a pipe segment identifier, and a flow direction identifier. Each drainage pumping station is abstracted as a pumping station node and associated with its inlet pipe segment and outlet pipe segment.

[0069] Based on the connectivity list, perform a reachability search along the flow direction for any two drainage pumping stations. When the pumping station node... There is at least one directed path along the flow direction that can reach the pump station node. At that time, and Determined to be an upstream and downstream related relationship and For upstream pumping stations, For downstream pumping stations, the upstream and downstream sets of each pumping station are written into the association cache within the scheduling cycle;

[0070] Based on the connectivity list, parallel relationships are identified. When the downstream node identifiers corresponding to the outlet pipe sections of at least two pumping stations are the same or the downstream pipe section identifiers are the same and the flow direction is consistent, the at least two pumping stations are identified as parallel drainage pumping stations and the parallel group identifier is recorded in the cache.

[0071] After obtaining the upstream and downstream correlations and parallel relationships, each drainage pumping station was analyzed according to the sampling time sequence. The inflow rate is calculated based on the sampling time sequence. The time alignment result is consistent with that of step S1 and satisfies The sampling time of the pump station was read from the processed monitoring data. Water level data of the inlet pool Combine the main outlet flow rate data and use it as the outlet flow rate. ;

[0072] Simultaneously, based on the as-built geometric parameters of the pump station's intake pool, a "correspondence between the intake pool level and effective volume" was established and solidified into a monotonically increasing level-effective volume calibration table. ,in For the first One calibrated liquid point and the unit is To and The corresponding effective volume and the unit is And use piecewise linear interpolation to Mapped to effective volume ,when When the liquid level exceeds the calibration range, The effective volume of the corresponding boundary liquid point is taken to ensure mapping stability;

[0073] Subsequently, the effective volume change rate is calculated based on the effective volume difference between adjacent sampling times and superimposed with the effluent flow rate to obtain the influent flow rate, according to the formula:

[0074] ;

[0075] in Indicates the sampling time The inflow rate is calculated and the unit is... Indicates the sampling time The outflow rate and the unit is Indicates the sampling time The effective volume of the inlet pool and the unit is Indicates the sampling time The effective volume of the inlet pool and the unit is Indicates the first Each sampling time and the unit is Indicates the first Each sampling time and the unit is Indicates the sampling period, with the unit being seconds (s).

[0076] right The first sampling time adopted As initialization to avoid missing The incalculable causes will be output along with the sequence of inflow calculations obtained by each pumping station in this scheduling cycle, as well as the upstream and downstream correlations and parallel relationships.

[0077] In this specific embodiment, S3 includes:

[0078] At the beginning of each scheduling cycle The execution first involves constructing a connection path index between pumping stations along the flow direction on the directed graph of the drainage network topology based on the upstream and downstream relationships. The connection path index is obtained by performing a reverse breadth-first search on each pumping station node to obtain the set of upstream pumping stations that it can reach and cache it to ensure that repeated queries do not result in a second traversal within the same scheduling cycle.

[0079] Subsequently, based on the "sewage treatment facility node" identifier in the pipeline topology, the target drainage pumping station connected to the sewage treatment facility is determined. The determination rule is as follows: starting from the sewage treatment facility node, all reachable pumping station nodes are searched in reverse along the topology, and the pumping station node with the shortest directed path length along the flow direction to the sewage treatment facility node and whose outlet pipe section directly merges into the sewage treatment facility inlet main pipe node is selected as the target drainage pumping station. This target drainage pumping station is uniquely identified as a first-level drainage pumping station and recorded as its level. ;

[0080] After identifying the target drainage pumping stations, a hierarchical ranking of the pumping stations is generated based on the connection path index, following a "step-by-step upstream expansion" approach. Specifically, all pumping stations directly connected to any... Pump station nodes that are connected to pump station nodes and have upstream and downstream relationships with them are assigned the value of secondary drainage pump stations and recorded as the level. Then direct all to any Pump station nodes that have merged into the system and have not yet been assigned a level are assigned the value of a Level 3 drainage pump station and recorded as such. This process continues until all pumping station nodes with upstream and downstream connections to the target drainage pumping station are assigned a unique level. and the level The pump station identifier is written into the hierarchical result table for subsequent scheduling and retrieval;

[0081] In terms of operational condition identification, rainfall data sequences are read from processed monitoring data. The cumulative rainfall is calculated in a continuous 30-minute time window, and the cumulative rainfall reaches a threshold. Time-based rainfall condition indicator Otherwise set Simultaneously, water level data of the receiving water body is read from the post-treatment monitoring data. and control elevation of each pump station outlet In comparison, among them The parameters are fixed in the parameter table based on the pump station design data. It was determined that the pumping station was affected by the backwater effect and a backwater condition indicator was installed. Otherwise set ;

[0082] Regarding the determination of control priorities, for each pumping station Based on the classification results, current liquid level safety status, downstream transport capacity occupancy, and inflow water characteristics indicated by water quality indicators, a priority score is generated for control. and will The global control priority is obtained by sorting from high to low. According to the formula:

[0083] ;

[0084] In the formula Indicates pumping station Control priority scoring and value range are Indicates the level weight and takes the value. This represents the safety weight of the liquid level and its value is... This indicates the weight of downstream transport capacity and its value is [value]. Represents water quality weight and takes values ​​of Represents the level factor and is determined by the pump station level. Determined according to the mapping table Take 1.00 at a time. Time to take Time to take Time to take This indicates the liquid level risk factor and is derived from the current inlet pool liquid level data. The preset alarm liquid level of the pump station The safety margin is determined as when When the value is 1.00, When taken as 0.80, When taking 0.50, When the value is 0.20, the value is determined by the top support condition indicator. When Set it directly to 1.00 to reflect the high risk caused by the top support. This indicates the downstream transport capacity occupancy factor, derived from the current outflow data of the main outlet pipe. Downstream transport capacity of the pumping station The ratio is determined as follows: 1.00 is taken when the ratio is not less than 0.90, and 1.00 is taken when the ratio is within the range of 0.90. When the ratio is less than 0.70, the value is taken as 0.30. The downstream control pipe section, determined by the pipeline network topology, is found in its design transport capacity table and remains unchanged during the scheduling cycle. This represents water quality factors and is derived from the chemical oxygen demand (COD) in the current water quality data. With threshold The comparison determines that when Use 1.00 if the concentration is high and 0.50 otherwise to reflect the impact of high-concentration influent on the stability of downstream treatment.

[0085] Among the parallel drainage pumping stations corresponding to the parallel relationship, the pumping stations within the parallel group are scored according to control priority. The scheduling order of the parallel group is obtained by sorting from high to low. When there are identical parallel groups within the group... Liquid level risk factor If the higher the level, the higher the priority; if they are still the same, the level factor will be used. Prioritize those with higher priority, and output the classification results of drainage pumping stations, the identification results of operating conditions, the global control priority, and the scheduling order of each parallel group.

[0086] In this specific embodiment, S4 includes:

[0087] At the beginning of each scheduling cycle The system executes and performs unified forecasting for all controlled drainage pumping stations, with the historical time window length set to [value missing]. There are 1 sampling point and the sampling period is 1 The prediction time domain length is set to One sampling point;

[0088] Based on the processed monitoring data, according to the pump station Capture historical time windows The sequence samples within are aligned according to the sampling time. The input feature vector at each sampling time. The data is obtained by splicing together data in the same order, representing the water level data of the inlet pool of the pumping station. Outflow data of main water pipe Water pump operating status data Water quality data Rainfall data Receiving water body water level data Rainfall condition indicators The pump station's top support condition indicator and the control priority score of the pumping station. ,in All values ​​are normalized values ​​after step S1. This indicates the sampling time of the pump station. The percentage of operating pumps obtained by summarizing the start and stop status of each water pump, and the result normalized according to the normalization parameter table in step S1. and The binary flag obtained from the working condition identification in step S3 remains unchanged within the same scheduling cycle and is copied and filled within the historical time window. The result calculated in step S3 Interval scoring and copying to fill within the historical time window;

[0089] A deep learning model, a spatiotemporal graph neural network model, is constructed and used for inference computation. This model uses drainage pumping stations as graph nodes and constructs an adjacency matrix based on upstream and downstream relationships. ,in The number of pump stations included in the scheduling is fixed at the number of pump station nodes in the topology data, and the adjacency matrix elements are... Set as pump station With pump station The value is set to 1 if there is a direct upstream / downstream relationship or a parallel relationship; otherwise, it is set to 0. A self-connection is added to each node to enable... ;

[0090] The spatial feature extraction module of the spatiotemporal graph neural network model adopts a two-layer graph convolutional network with the number of hidden channels set to 32 and 64 respectively, the activation function set to ReLU and the inactivation rate set to 0.1, so as to perform graph convolutional aggregation on the input features of adjacent pump station nodes at each sampling time.

[0091] The time-series modeling module of the spatiotemporal graph neural network model employs a two-layer gated recurrent unit network and uses the 64-dimensional node embedding sequence output by the aforementioned graph convolution as input. The hidden state dimension of the gated recurrent unit network is set to 64 and is arranged in time steps from... to The timing context is obtained through recursion;

[0092] The output head of the spatiotemporal graph neural network model uses three sets of independent fully connected layers and maps the terminal hidden state of the time series modeling module to a sequence of predicted inflow values ​​in the prediction time domain for each pump station node. Predicted water level sequence of the inlet pool and water quality prediction series ,in All are normalized values ​​consistent with step S1, and after output, they are denormalized into engineering quantities according to the normalization parameter table for use in step S5;

[0093] After obtaining the predicted liquid level sequence of the inlet tank, an overflow risk prediction sequence is generated. The overflow risk prediction value is calculated using the formula:

[0094] ;

[0095] in Indicates pumping station At the predicted time The predicted value of overflow risk and the range of values ​​are Indicates pumping station At the predicted time The reverse normalized predicted influent level and the unit is Indicates pumping station The preset alarm liquid level, in meters, is fixed and given by the pump station parameter table. Indicates the normalized span of liquid level risk and sets it to Used to map the magnitude of exceeding the alarm level to a risk increment. Represents any prediction time within the prediction time domain that satisfies and Finally, the predicted inflow rate, the predicted inlet pool level, the predicted water quality, and the predicted overflow risk are output as the prediction results.

[0096] In this specific embodiment, S5 includes:

[0097] At the beginning of each scheduling cycle Execute and receive inflow forecast values Predicted water level in the inlet tank Water quality prediction values and overflow risk prediction value Simultaneously receive global control priority and target drainage pumping station identifier. Read the current water level data in the inlet tank from the processed monitoring data. Current outflow data for the main water pipe The current operating status data of the water pumps is used to determine the current operating conditions of each drainage pumping station;

[0098] The set of prediction times in the prediction time domain is set as follows: and in the same way as step S4 and As a fixed prediction step size and prediction number;

[0099] The downstream conveying capacity of each drainage pumping station is determined based on the pipeline topology and design parameter table. The determination rule is to search along the topological flow direction for the unique main path between the outlet of the pumping station and the inlet of its nearest downstream pumping station or sewage treatment facility, and take the minimum design conveying capacity of all pipe segments on this main path as the downstream conveying capacity of the pumping station. The designed transport capacity of the pipeline segment is fixed and stored in the design parameter table by parameters such as pipe diameter, slope and roughness, and remains unchanged during the scheduling cycle;

[0100] The scheduling constraints are constructed and enforced during the solution process. These constraints include ensuring that the actual operating liquid level of each drainage pumping station in the prediction time domain does not exceed the preset alarm liquid level. The total outflow from each drainage pumping station within the predicted time period shall not exceed the corresponding downstream conveying capacity. The control decisions are determined sequentially from high to low according to the control priorities given in step S3, and sequentially according to the scheduling order within the parallel group among the parallel drainage pumping stations.

[0101] Under the premise of meeting the above constraints, for each drainage pumping station Candidate control schemes are generated and solved using a discrete enumeration method. Each candidate control scheme consists of pump start-stop control parameters and variable frequency drive (VFD) operation setpoints. The pump start-stop control parameters include the pump start-up liquid level threshold. With the water pump stop liquid level threshold The variable frequency operation setting value is the unified variable frequency setting value within the scheduling cycle of this pumping station. ;

[0102] The set of candidate control schemes is fixed in this embodiment as follows: Take in sequence , And mandatory satisfaction To avoid frequent start-stop cycles, Take in sequence ;

[0103] For any candidate control scheme, its impact on the fluctuation range of liquid level, effluent from each pumping station, and effluent flow rate of the target drainage pumping station needs to be evaluated within the prediction time domain. During the evaluation, a network hydraulic calculation model is established, and the candidate control scheme is input as boundary conditions into the model to obtain the total effluent flow rate sequence of the target drainage pumping station. The network hydraulic calculation model adopts a one-dimensional unsteady flow dynamic wave model and constructs a computational network based on the network topology. Network elements include pipe segments and nodes. Pipe segment parameters include length, diameter, slope, and Manning roughness, taken from the design parameter table. Node parameters include manhole bottom elevation and storage capacity, taken from as-built data. The numerical solution uses an implicit difference scheme and progresses in the prediction time domain with a hydraulic step size of 10 s. The model boundary conditions include: the inflow boundary of each pumping station node is taken as the predicted inflow rate. The water level of the receiving water body in the processed monitoring data is at the boundary of the receiving water body. The observed values ​​at time 1 and 2 remain constant within the prediction time domain. The outflow boundary of the pumping station is determined by the candidate control scheme and applied according to the following logic: when the liquid level in the pumping station's inlet pool reaches or exceeds 100 at the simulated time 2, the boundary is determined by the candidate control scheme and applied according to the following logic: When the pumping station is put into operation under the same conditions as the current pumping station, the pumping units will be merged, and the flow rate of each operating pump will be set according to the rated flow rate and frequency settings. The outflow from the pumping station at that moment was obtained by converting the proportional relationship. When the water level in the pumping station's inlet pool drops to [a certain value] at the simulated moment... The following steps involve setting all pumps at the pumping station to shutdown and zeroing the outflow from the pumping station. This allows for the generation of a sequence of total outflow rates for each pumping station in the hydraulic model that matches the candidate control scheme, and ultimately, the generation of the total outflow rate for the target drainage pumping station. ;

[0104] In obtaining The fluctuation range of the outflow rate of the target drainage pumping station is then calculated and a hard constraint is applied. The hard constraint is set so that the fluctuation range of the outflow rate of the target drainage pumping station in the prediction time domain does not exceed a preset threshold. The fluctuation range of the outflow rate is determined by the predicted time domain. The difference between the maximum and minimum values ​​is used to determine this.

[0105] From the set of candidate control schemes that satisfy the constraints of liquid level, downstream transport capacity, priority order, and hard constraint of effluent fluctuation amplitude, the scheduling target evaluation value is calculated, and the candidate control scheme with the optimal scheduling target evaluation value is selected as the scheduling control command. The scheduling target evaluation value is calculated according to the formula:

[0106] ;

[0107] In the formula This represents a set of candidate control scheme combinations covering all drainage pumping stations. This represents the scheduling target evaluation value of the candidate control scheme combination, with a smaller value indicating better performance. Indicates the overflow risk weight and takes Indicates energy consumption weight and takes Indicates the weight of water flow fluctuation and takes Indicates pumping station At the predicted time The predicted value of overflow risk, This indicates that all pumping stations With all prediction times within the prediction time domain Take the maximum value to reflect the global worst-case overflow risk. Indicates the combination of candidate control schemes The energy consumption index in the prediction time domain, in kWh, is obtained by converting the power of the number of pumps operating at each pumping station at each time point and the frequency setpoint output by the hydraulic model, and then integrating over time. Indicates the target drainage pumping station At the predicted time The total outflow rate is output by the hydraulic calculation model of the pipe network. and These represent all prediction times within the prediction time domain. Take the maximum and minimum values;

[0108] During the solution process, a greedy enumeration strategy with a fixed order is used to construct the solution according to the control priority from high to low. That is, after determining the candidate control schemes for high-priority pumping stations, these schemes are used as fixed boundary conditions in subsequent hydraulic calculations, and the process continues to enumerate and filter for the next pumping station, resulting in the final... Simultaneously satisfying control sequence constraints and global hydraulic constraints;

[0109] Combine the selected optimal candidate control schemes The water pump start-up liquid level threshold corresponding to each drainage pumping station. Water pump stop liquid level threshold With variable frequency operation set value Together they form the scheduling and control instructions.

[0110] In this specific embodiment, S6 includes:

[0111] After the dispatch control command is output in step S5, it is executed immediately and sent to the corresponding automatic control system of each drainage pumping station one by one. The dispatch control command is applied to each pumping station. Fixed water pump start-up liquid level threshold Water pump stop liquid level threshold With variable frequency operation set value It also includes an instruction serial number, target pump station identifier, and check code to ensure traceability and prevent crosstalk;

[0112] The dispatching platform and the automatic control systems of each pumping station use OPC UA over TCP for parameter transmission. and Write the start / stop threshold register of the pump station PLC and After writing the variable frequency setpoint register, the PLC sets the parameter validity flag to valid and sends the corresponding register mirror value back to the dispatch platform. When the dispatch platform receives the feedback and the mirror value matches the sent value, it determines the sending is successful and records the sending time. ,in This serves as a unified timestamp for the scheduling platform when it successfully confirms the completion of an instruction.

[0113] The pump station PLC uses its local control logic to... and The water pump is started and stopped based on the water level threshold of the inlet tank. The frequency setting is sent to the frequency converter via fieldbus to control the variable frequency operation of the water pumps. Simultaneously, the PLC continuously uploads water pump operating status data to reflect changes in the start / stop status of each pump. After the command is issued, the dispatch platform continuously monitors the water pump operating status data of the pumping station and records the moment of the first change. ,in For relative to The unified timestamp when the first change in the water pump operating status data was detected;

[0114] The scheduling platform is based on and Jointly determine the effective time of the scheduling control command The effective time is determined by the formula:

[0115] ;

[0116] in Indicates pumping station The time when the scheduling and control command takes effect. Indicates pumping station The moment when the water pump operating status data first changes Indicates pumping station The moment when the dispatch control command is successfully issued is confirmed. This represents the maximum time to wait for changes in the pump's operating status, and takes... Indicates the pump station index;

[0117] since The system then enters the feedback collection phase and continuously collects data on the actual inlet water level, actual outlet main flow rate, actual pump operating status, and actual water quality of the pump station within a preset feedback sampling period. The feedback sampling interval is set to [specified value]. And the duration of feedback collection is set to At each feedback sampling time, the scheduling platform adds a unified timestamp to the above four types of actual data and writes them into the same record to form execution feedback monitoring data. Each record of the execution feedback monitoring data includes at least the pump station identifier, sampling time, actual inlet pool liquid level, actual outlet main flow rate, actual combination of operating pumps, and actual water quality indicators.

[0118] In this specific embodiment, S7 includes:

[0119] The feedback monitoring data collection should be performed immediately after completion, and the feedback monitoring data should cover each drainage pumping station. From the time its instruction takes effect The feedback sampling interval and the sampling interval are It includes actual water level data in the inlet tank. Actual outflow data of the main water pipe Actual water pump operating status data and actual water quality data ,in Indicates the first Each feedback sampling moment;

[0120] For each drainage pumping station Traverse in chronological order The preset alarm liquid level of the pump station When the same pumping station node meets the requirements and When Record the start time of the overflow event when the conditions are met. and When The overflow event end time is recorded, and an overflow event set is obtained by pairing the start time with the end time one by one. The overflow count is counted according to the number of elements in the event set, and the overflow duration is obtained by summing the time difference between the end time and the start time of each event, and is calculated when the end of the feedback sampling interval is still in the range. In the state, the end time of the feedback sampling interval is taken as the end time of the event;

[0121] Secondly, within the feedback sampling interval, for each drainage pumping station Pick The maximum value and with The safety margin of the critical node liquid level is obtained by comparison. The safety margin of the critical node liquid level is defined as follows: The difference from the maximum value is set to 0 when the difference is negative to indicate that there is no safety margin;

[0122] Targeting the drainage pumping station again Within the feedback sampling interval The fluctuation range of the outflow rate of the target drainage pumping station is obtained by taking the maximum and minimum values ​​and calculating the difference.

[0123] Meanwhile, the running time of each pump is calculated based on the actual operating status data. The running time calculation rule is to identify the change time from pump stop to pump start in the start-up sequence of each pump as the start time of operation and identify the change time from pump start to pump stop as the end time of operation. The running time is the sum of the lengths of all operating intervals. The energy consumption index is obtained by multiplying the running time of each pump by the rated power of the pump and summing them up by pump station. The rated power of the pump is fixed by the equipment ledger parameter table and uniquely indexed by the pump identifier.

[0124] Subsequently, the number of overflows, overflow duration, critical level safety margin, fluctuation range of the target drainage pumping station's outflow, and energy consumption indicators are compared with preset targets to generate performance evaluation results. In this embodiment, the preset target is fixed as the number of overflows not exceeding [a certain threshold]. The overflow duration is no greater than The safety margin of liquid level at any critical node of the pumping station shall not be less than The fluctuation range of the outflow rate of the target drainage pumping station shall not exceed [a certain value]. And the energy consumption index is not greater than the energy consumption index calculated using the same statistical rules in the previous scheduling cycle;

[0125] When the performance evaluation results fail to meet the preset targets, the execution model self-updates and incorporates the current execution feedback monitoring data into the supervision sample. Specifically, the execution feedback monitoring data is preprocessed again through time alignment, missing value handling, outlier removal, and normalization to obtain updated processed monitoring data. The inflow rate is then calculated based on the updated inlet tank level data and outlet main pipe flow rate data according to the inflow rate calculation rules. As a label for monitoring inflow, it also includes actual inflow tank level data. Compared with actual water quality data As a liquid level monitoring label and a water quality monitoring label;

[0126] Then, using the same historical time window length as step S4... With the predicted time domain length Construct supervised sample pairs, where the input of each supervised sample pair is the historical time window sequence features organized by the pump station, including the same feature composition as in step S4, operating condition indicators and control priority scores, and using the same adjacency matrix as in step S4. Describes the topological relationships of the pumping stations; the output of each supervised sample pair is the corresponding prediction time domain. and ;

[0127] Based on this, the spatiotemporal graph neural network model is retrained to update the model parameters. The retraining adopts an incremental training method based on the current model parameters, and the training hyperparameters are fixed as the Adam optimizer and the learning rate. Weight decay coefficient Batch size Number of training rounds The supervised samples are divided into training and validation sets in chronological order, with the validation set accounting for the largest proportion of the samples. Over time, the loss function for retraining is calculated using the following formula:

[0128] ;

[0129] in This represents the training loss value. Indicates the inflow loss weight and takes Indicates the weight of liquid level loss and takes Indicates the weight of water quality loss and takes This represents the calculation of the mean square error between the predicted value and the monitoring label across all pump station nodes and all prediction steps. This represents the predicted inflow rate output by the model. Indicates the supervision label for the inflow calculation value. This represents the predicted water level in the inlet tank output by the model. This label indicates the actual water level data monitoring data of the inlet pool. This represents the water quality prediction value output by the model. This indicates the actual water quality data monitoring label, and completion. After each training round, the model parameters with the minimum loss on the validation set are selected as the updated model parameters and used in the deep learning prediction step of the next scheduling cycle.

[0130] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0131] This invention forms a closed-loop influence chain for the scheduling decisions of drainage pumping stations through a combination of algorithms including "data preprocessing and unified time benchmark, topology coupling modeling, short-term prediction, constraint optimization solution, command issuance and feedback evaluation". Specifically, it first calculates the inflow rate based on the relationship between the inlet tank level and effective volume and the flow rate of the outlet main pipe, obtaining a unified input that can be used for learning; then, it identifies upstream and downstream and parallel relationships based on the pipeline topology, classifies pumping stations and generates control priorities, transforming the scheduling objective from local response of a single station to multi-pumping station collaboration; subsequently, it uses a deep learning model to output the prediction results of inflow rate, liquid level, water quality and overflow risk in the prediction time domain, and solves the pump start-stop threshold and frequency conversion setpoint under liquid level alarm constraints, downstream transport capacity constraints and priority order constraints, while setting threshold constraints for the effluent flow fluctuation of the target pumping station, thereby achieving the technical effects of controllable water level, reduced overflow, stable influent to sewage treatment facilities and controllable energy consumption under rainy and sunny conditions.

[0132] This invention addresses the technical problem of "difficulty in comprehensively characterizing the coupling of multiple pumping stations and external influences" by making scenario-oriented improvements: First, it introduces a spatiotemporal modeling approach consistent with the pipeline network topology, using the adjacency relationship of pumping stations as the spatial structure input to the deep learning model. This allows upstream water inflow, parallel inflow, and backwater effects to be explicitly propagated during feature aggregation, thereby improving the accuracy of short-term prediction and risk assessment. Second, it couples overflow risk prediction with scheduling optimization, adopting an evaluation and constraint system centered on overflow risk, energy consumption, and fluctuations in the outflow of the target pumping station. It can also combine the pipeline network hydraulic calculation model to evaluate the feasibility and effectiveness of candidate schemes, making the generated control parameters more consistent with engineering constraints. Third, it constructs a performance evaluation and retraining mechanism based on execution feedback, which can continuously update the model and strategy according to seasonal infiltration changes, rainfall pattern changes, and equipment status changes, reducing reliance on human experience and enhancing long-term operational stability.

Claims

1. A deep learning-based method for scheduling drainage pumping stations, comprising: S1. Within the preset scheduling cycle, collect the operation monitoring data, external impact data and topological relationship of drainage pipe network of each drainage pumping station, add timestamps and perform preprocessing to obtain the processed monitoring data. S2. Determine the upstream and downstream relationships and parallel relationships of each drainage pumping station based on the pipeline network topology. Calculate the inflow rate of each drainage pumping station based on the influent tank level data and the outflow rate data in the treated monitoring data, combined with the correspondence between the influent tank level and the effective volume. S3. Classify the drainage pumping stations according to the upstream and downstream relationships to obtain the classification results. Identify the operating conditions based on the treated monitoring data to obtain the operating condition identification results, determine the control priority of the drainage pumping stations, and determine the target drainage pumping stations connected to the sewage treatment facilities based on the pipeline network topology. S4. Extract the temporal features within a preset historical time window from the processed monitoring data, and input them into a deep learning model along with the operating condition identification results and the control priority of the drainage pumping station. Output the prediction results within the preset prediction time domain. S5. Based on the prediction results, the control priority of the drainage pumping station, and the current inlet tank level data, current outlet main flow data, and current pump operating status data from the processed monitoring data, establish scheduling constraints. The scheduling objectives are to reduce overflow risk, reduce energy consumption, and ensure that the fluctuation range of the outlet flow of the target drainage pumping station does not exceed a preset threshold. Solve for the pump start-stop control parameters and variable frequency operation setpoints of each drainage pumping station, and combine the pump start-stop control parameters and variable frequency operation setpoints into scheduling control commands. S6. Send dispatch control commands to the automatic control system of the corresponding drainage pumping station to control the operation of the pumps, and collect execution feedback monitoring data after the commands are executed; S7. Based on the execution feedback monitoring data, the performance evaluation results are obtained. When the results do not meet the preset goals, the supervision samples are updated based on the execution feedback monitoring data. The deep learning model is then retrained using the updated supervision samples to update the model parameters, so that the updated model parameters can be used in the next scheduling cycle.

2. The deep learning-based drainage pumping station scheduling method according to claim 1, S1 includes: Within the preset scheduling cycle, operation monitoring data, external impact data, and topological relationships of drainage pipe networks are collected for each drainage pumping station. The operation monitoring data includes inlet pool level data, outlet main pipe flow data, pump operation status data, and water quality data. The external impact data includes rainfall data and receiving water body level data. The collected operational monitoring data and external impact data are each appended with a timestamp, and the timestamps are converted to a unified time base. Based on a preset sampling period, time alignment is performed on the operation monitoring data and the external influence data after the additional timestamps are applied to obtain aligned time-series data; The aligned time-series data are sequentially processed by missing value handling, outlier removal, and numerical normalization to obtain the processed monitoring data.

3. The deep learning-based drainage pumping station scheduling method according to claim 1, S2 includes: Based on the pipeline network topology, the upstream and downstream relationships and parallel relationships of each drainage pumping station are determined. The upstream and downstream relationships include the connectivity between each drainage pumping station along the flow direction of the drainage pipeline network, and the parallel relationships include the connectivity between at least two drainage pumping stations that converge into the same drainage pumping station or the same drainage pipeline. Based on the inlet pool level data in the processed monitoring data, the inlet pool level at each time is converted into the corresponding effective volume using the correspondence between the inlet pool level and the effective volume, and the effective volume change rate is calculated based on the effective volume difference between adjacent time points and the time difference between adjacent time points. The outflow rate at each time point is determined based on the outflow rate data of the main outflow pipe in the processed monitoring data; The inflow rate of each drainage pumping station is calculated based on the outflow rate and the effective volume change rate, wherein the inflow rate is equal to the sum of the outflow rate and the effective volume change rate.

4. The deep learning-based drainage pumping station scheduling method according to claim 1, S3 includes: Based on the upstream and downstream relationships, the connection path between each drainage pumping station along the drainage network is determined, and based on the network topology, the drainage pumping station connected to the sewage treatment facility is determined as the target drainage pumping station. The target drainage pumping station is identified as a first-level drainage pumping station, and the drainage pumping stations that have the upstream and downstream relationship with the first-level drainage pumping station and flow into the first-level drainage pumping station along the connection path are identified as second-level drainage pumping stations. The process continues to extend upstream along the connection path to obtain the drainage pumping station classification results. Based on the rainfall data in the processed monitoring data, it is determined whether a rainfall process has occurred, and based on the water level data of the receiving water body and the liquid level data of the inlet pool in the processed monitoring data, it is determined whether there is an impact from the backwater of the receiving water body, thus obtaining the operating condition identification result. Based on the classification results of the drainage pumping stations and combined with the inlet pool level data, outlet main flow data, pump operation status data and water quality data in the processed monitoring data, a corresponding control priority score is generated for each drainage pumping station, and the control priority of the drainage pumping station is determined according to the control priority score. The control priority score is at least related to the drainage pumping station level, the safety margin of the inlet pool level from the preset alarm level, the degree to which the outlet main flow is close to the downstream transport capacity, and the nature of the incoming water indicated by the water quality data. And among the parallel drainage pumping stations corresponding to the parallel relationship, the scheduling order of the parallel drainage pumping stations is determined according to the control priority of the drainage pumping stations.

5. The deep learning-based drainage pumping station scheduling method according to claim 1, S4 includes: The processed monitoring data is used to extract sequence samples within a preset historical time window for each drainage pumping station. The sequence samples include, at each sampling time, the inlet pool level data, the outlet main flow rate data, the pump operating status data, the water quality data, the rainfall data, and the receiving water level data for that drainage pumping station. The sequence samples, the operating condition identification results, and the drainage pump station control priority are jointly encoded into the input sequence of the deep learning model, and the deep learning model outputs the inflow prediction value, the inlet pool liquid level prediction value, and the water quality prediction value within the preset prediction time domain. An overflow risk prediction value is generated based on the comparison between the predicted water level in the inlet tank and the preset alarm level within the preset prediction time domain. When the predicted water level in the inlet tank reaches or exceeds the preset alarm level, the overflow risk prediction value at the corresponding time is increased.

6. The deep learning-based drainage pumping station scheduling method according to claim 1, S5 includes: Based on the predicted values ​​of the inlet pool level and overflow risk in the prediction results, the level risk of each drainage pumping station within the preset prediction time domain is determined. Based on the current inlet water level data, current outlet main flow data, and current pump operating status data in the processed monitoring data, the current operating conditions of each drainage pumping station are determined. Based on the pipeline topology, the downstream transport capacity of each drainage pumping station is determined. This downstream transport capacity, along with the liquid level risk level and the control priority of the drainage pumping stations, is used to construct the scheduling constraints. These constraints include: the liquid level at the corresponding node of each drainage pumping station does not exceed the preset alarm liquid level; the total outflow of each drainage pumping station does not exceed its corresponding downstream transport capacity; and a control sequence that satisfies the control priority of the drainage pumping stations. Under the premise of satisfying the scheduling constraints, candidate control schemes are generated for each drainage pumping station. Each candidate control scheme includes the pump start-stop control parameters and variable frequency drive (VFD) operation setpoints for that drainage pumping station. The pump start-stop control parameters include a pump start liquid level threshold and a pump stop liquid level threshold. For each candidate... The control scheme calculates a scheduling target evaluation value, which is at least related to the overflow risk prediction value, energy consumption, and the fluctuation range of the outflow of the target drainage pumping station. The fluctuation range of the outflow of the target drainage pumping station is determined by the difference between the maximum and minimum values ​​of the outflow of the target drainage pumping station's main outlet pipe corresponding to the candidate control scheme within the preset prediction time domain, and the fluctuation range of the outflow of the target drainage pumping station is required not to exceed the preset threshold. From the candidate control schemes that meet the scheduling constraints and satisfy the requirement that the fluctuation range of the outflow of the target drainage pumping station does not exceed the preset threshold, the candidate control scheme with the optimal scheduling target evaluation value is selected, and the pump start-stop control parameters and variable frequency operation setpoints corresponding to the candidate control scheme are determined as the scheduling control command.

7. The deep learning-based drainage pumping station scheduling method according to claim 1, S6 includes: The dispatch control command is sent to the automatic control system of the corresponding drainage pumping station one by one. The automatic control system controls the start and stop of the pumps according to the pump start and stop control parameters in the dispatch control command, and controls the pumps to operate at variable frequency according to the variable frequency operation set value in the dispatch control command. After the dispatch control command is sent, the time of issuance of the dispatch control command and the time of change of the pump operation status data are recorded, and the effective time of the dispatch control command is determined by the time of issuance and the time of change. From the effective time, the actual inlet water level data, the actual outlet main flow rate data, the actual water pump operating status data, and the actual water quality data are continuously collected within the preset feedback sampling period. Timestamps are added to the actual inlet water level data, the actual outlet main flow rate data, the actual water pump operating status data, and the actual water quality data to obtain the execution feedback monitoring data.

8. The deep learning-based drainage pumping station scheduling method according to claim 1, S7 includes: An overflow event is determined based on the comparison between the actual inlet water level data and the preset alarm level in the execution feedback monitoring data. Specifically, at the corresponding node of the same drainage pump station, the overflow event is defined as the start time when the actual inlet water level data changes from less than the preset alarm level to reaching or exceeding the preset alarm level, and the overflow event is defined as the end time when the actual inlet water level data changes from reaching or exceeding the preset alarm level to less than the preset alarm level. The number of overflows is counted based on the start and end times of each overflow event, and the duration of each overflow event is calculated and accumulated to obtain the overflow duration. Within the preset feedback sampling period, the safety margin of the critical node liquid level is calculated based on the difference between the maximum value of the actual inlet pool liquid level data at the corresponding node of each drainage pumping station in the execution feedback monitoring data and the preset alarm liquid level. Within the preset feedback sampling period, the fluctuation range of the outflow of the target drainage pumping station is calculated based on the difference between the maximum and minimum values ​​of the actual outflow of the target drainage pumping station in the execution feedback monitoring data. Within the preset feedback sampling period, the running time of each pump is determined based on the actual pump operating status data in the execution feedback monitoring data, and the energy consumption index is calculated based on the pump rated power and the running time. The performance evaluation results are obtained by comparing the number of overflows, the duration of overflows, the safety margin of the liquid level at key nodes, the fluctuation range of the outflow of the target drainage pumping station, and the energy consumption indicators with the preset targets. When the performance evaluation results do not meet the preset targets, the execution feedback monitoring data is preprocessed to update the processed monitoring data, and the inflow calculation value is calculated based on the inlet pool level data and the outlet main pipe flow data in the processed monitoring data. The monitoring samples are updated based on the updated post-processing monitoring data, inflow calculation value, actual influent pool level data, and actual water quality data. The updated monitoring samples are then used to retrain the deep learning model to update the model parameters, so that the updated model parameters can be used in the deep learning prediction step of the next scheduling cycle.

9. A deep learning-based drainage pumping station scheduling method according to claim 5, characterized in that, The deep learning model is a spatiotemporal graph neural network model. Based on the pipeline network topology, it constructs an adjacency matrix or graph edge set of pump station nodes. After performing graph convolution feature aggregation on the sequence samples of adjacent pump station nodes, it performs time series modeling and outputs the predicted inflow rate, the predicted water level in the inlet pool, and the predicted water quality within the preset prediction time domain.

10. A deep learning-based drainage pumping station scheduling method according to claim 6, characterized in that, Calculating the total outflow of the target drainage pumping station for each candidate control scheme includes: establishing a hydraulic calculation model of the pipeline network based on the pipeline network topology, and inputting the pump operating status and frequency conversion operation setpoint corresponding to the candidate control scheme as boundary conditions into the hydraulic calculation model of the pipeline network to obtain the sequence of total outflow of the target drainage pumping station in the preset prediction time domain.

Images