Multi-factor coupled rural sewage load feedforward prediction intelligent control method and system

Abstract

This invention provides a multi-factor coupled intelligent control method and system for feedforward prediction of rural sewage load, belonging to the field of intelligent control technology. The invention employs a lightweight, time-series shrinking adaptive LSTM base model pre-trained offline using historical data collected from stations throughout the year in the cloud, generating a base model containing long-term annual sewage discharge patterns and distributing it to the edge gateway. At the edge, only the samples of the last 7 days are continuously cached, the base model is loaded, and the backbone network is frozen. Only the adaptive correction layer is enabled to perform daily incremental self-learning, while simultaneously integrating four-dimensional rural-specific exogenous fine-tuning factors to achieve advance prediction of influent COD peak, ammonia nitrogen peak, and sewage treatment volume for the next 2 and 4 hours. Feedforward pre-regulation is performed based on the predicted load level, and combined with real-time effluent water quality feedback correction to form a dual closed-loop control. This invention significantly reduces edge hardware costs, supports independent offline operation when the network is down, and is suitable for unattended, precise management scenarios of decentralized rural sewage treatment plants.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to a multi-factor coupled intelligent control method and system for feedforward prediction of rural sewage load. Background Technology

[0002] Rural decentralized sewage treatment plants face significantly different operational challenges compared to centralized urban sewage treatment plants due to their dispersed layout, small scale, and lack of dedicated maintenance personnel. These challenges manifest primarily in the following ways: rural domestic sewage discharge exhibits typical non-stationary, pulse-like characteristics, influenced by multiple factors such as villagers' daily routines, holiday travel rushes, rural catering activities, and rainfall. This results in drastic fluctuations in influent water quality and quantity within short periods, posing a severe test to the stable operation of the treatment system.

[0003] Currently, the automation control of decentralized rural wastewater treatment plants, both domestically and internationally, generally adopts a closed-loop control scheme based on real-time sensor feedback. The typical approach involves continuously collecting real-time parameters such as dissolved oxygen in the biological treatment tank, effluent ammonia nitrogen, effluent chemical oxygen demand, and influent flow rate through online sensors. Based on preset proportional-integral-derivative control thresholds or segmented logic, the operating frequency of the aeration fan and the start / stop and dosing frequency of the chemical dosing pump are adjusted with a lag when water quality parameters deviate from the set range. This type of control logic is a passive response mode; adjustments are only triggered after an abnormality in water quality occurs.

[0004] Under the aforementioned control scheme, the following technical problems exist: When a shock load on rural sewage has already occurred, the sensors collect water quality data that has already entered the biological treatment system or reached the effluent outlet. If equipment adjustments are initiated at this point, there is a lag of several hours in the adaptation of the bacterial community and the adjustment of process parameters in the biological treatment system. In scenarios where rural areas experience peak sewage discharge due to concentrated washing and laundry activities in the morning and evening, or when sewage discharge increases exponentially due to returning residents on weekends and holidays, the shock load is intense and concentrated in time. Passive, delayed control is insufficient to absorb the load shock in a short period, leading to short-term exceedances of effluent water quality and drastic fluctuations in water quality. Therefore, the existing control logic inherently suffers from lag and insufficient shock load resistance when dealing with the uniquely strong shock loads of rural sewage.

[0005] In terms of predictive modeling, existing methods for forecasting wastewater quality and quantity mostly rely on single-type time-series data of water quality and flow collected by instruments within the station to build predictive models. These methods attempt to extract discharge patterns from the station's own historical operational data. However, fluctuations in rural wastewater discharge cannot be fully characterized by station parameters; they are substantially related to external social and environmental factors such as the village's resident population, holiday periods, weather and rainfall, and customer flow from rural tourism and catering businesses. Existing technical solutions do not incorporate these rural-specific disturbance factors during modeling, resulting in a single source of information. This leads to a significant drop in prediction accuracy when data volume is low or when discharge patterns temporarily change, and insufficient adaptability and generalization ability for different scenarios.

[0006] Regarding computing power deployment architecture, existing technical solutions exhibit two typical modes. The first is a pure edge solution, where all historical data storage, model training, and predictive inference are deployed on-site edge gateways. This solution requires edge hardware with the storage and computational capabilities to handle a full year's worth of time-series data. The storage capacity and computing power of low-end embedded gateways are insufficient, leading to excessively long training times or ineffective model convergence. The second is a pure cloud solution, where all prediction and control computational tasks are completed in real-time in the cloud. This solution heavily relies on real-time network connectivity. However, the widely deployed 4G and narrowband IoT networks in rural areas suffer from weak signal coverage, low bandwidth, and intermittent network outages. Cloud transmission latency is high, data packet loss is frequent, and on-site equipment completely loses its control capabilities during outages, making real-time on-site management impossible.

[0007] Regarding control strategies, existing technologies mostly employ fixed-level or single-threshold control methods, failing to match process parameters in advance based on actual load change trends. During low-load periods, equipment may still operate at higher parameters, resulting in ineffective consumption of electricity and chemicals; during periods of high-impact loads, the equipment's control capabilities are insufficient to cope with the risk of exceeding standards. Therefore, how to achieve a dynamic balance between ensuring stable effluent compliance and energy conservation and emission reduction is also a challenge faced by existing technologies.

[0008] In summary, existing technologies in the field of intelligent control of rural decentralized sewage treatment plants suffer from technical problems such as control lag, limited modeling factors, mismatch between computing power deployment and network conditions, and crude control strategies. There is still a lack of a technical solution that can effectively cope with strong shock load fluctuations in rural sewage under low hardware costs and weak network conditions, and achieve stable effluent compliance with standards while also saving energy and reducing consumption. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a multi-factor coupled intelligent control method and system for feedforward prediction of rural sewage load. This system aims to achieve advanced prediction of influent load for the next 2 to 4 hours and feedforward pre-regulation based on load levels by constructing a lightweight, time-series shrinking adaptive LSTM basic model pre-trained with full-scale historical data in the cloud, and a cloud-edge layered computing power decoupling architecture that integrates only four-dimensional rural exogenous factors (holidays, weather, population, and business passenger flow) on the edge side for 7-day small-sample incremental self-learning. This is supplemented by real-time effluent quality feedback for micro-correction, forming a dual closed-loop control. Thus, under the conditions of adapting to weak rural networks or even independent operation during network outages and significantly reducing edge hardware costs, it fundamentally eliminates shock-induced effluent exceeding standards and achieves significant energy savings and consumption reduction.

[0010] To achieve the above objectives, this application proposes a multi-factor coupled intelligent control method for feedforward prediction of rural sewage load, which is used for the automated control of rural decentralized sewage treatment plants, including cloud pre-training steps and edge local operation steps. The cloud-based pre-training steps include: Obtain historical time-series data and static external source information of rural decentralized sewage treatment plants over the past year; The historical time-series data and the static external information of the site are preprocessed, and a lightweight time-series shrinkage adaptive LSTM basic model is trained offline based on the preprocessed historical time-series data to obtain a basic model containing the long-term annual sewage discharge pattern. The basic model is then distributed to the edge gateway of the corresponding site. The lightweight time-series shrinkage adaptive LSTM basic model includes a time-series shrinkage module, a lightweight LSTM memory layer, and an adaptive correction layer. The edge local operation steps include: Real-time collection of internal working condition data, and simultaneous acquisition of rural-specific external fine-tuning factor data; The intrinsic working condition data and the rural-specific extrinsic fine-tuning factor data are preprocessed to construct a multi-source fusion time-series feature matrix; The edge gateway loads the base model, freezes the network parameters of the time-series shrinking module and the lightweight LSTM memory layer, enables only the adaptive correction layer, and uses locally stored historical running samples to perform incremental self-learning on the adaptive correction layer to obtain an updated model adapted to local sewage discharge conditions. The multi-source fusion time-series feature matrix is ​​input into the update model to predict the wastewater treatment load for the next 2 and 4 hours. Based on the predicted wastewater treatment load, feedforward pre-control is implemented in advance, and feedback correction is performed in conjunction with real-time effluent water quality parameters, forming a dual closed-loop control of feedforward prediction and feedback correction.

[0011] As a further solution, the cloud pre-training step also includes: every 12 months, summarizing the newly added historical time-series data of the site throughout the year, re-executing the preprocessing and offline training to obtain an updated base model, and remotely distributing the updated base model to the edge gateway of the corresponding site.

[0012] As a further solution, the edge gateway only keeps a rolling record of local operating samples for the past 7 days, and the incremental self-learning, the prediction of wastewater treatment load, and the feedforward pre-regulation are all performed locally on the edge gateway. When the network connection is normal, the edge gateway will synchronously upload local operating data and samples to the cloud for archiving; When the network is interrupted, the edge gateway independently completes prediction and control based on local cached data, and resumes the transmission of cached data after the network is restored. As a further solution, the adaptive correction layer of the base model adopts an incremental self-learning mechanism based on small samples from the past 7 days. Every 24 hours, it reads the running samples from the past 7 days in the local cache for lightweight iterative fine-tuning, automatically adapting to the sewage discharge characteristics, seasonal water temperature and business changes of different villages.

[0013] As a further solution, the wastewater treatment load includes at least the wastewater treatment volume, peak influent COD, peak ammonia nitrogen, and maximum instantaneous flow rate; when the time-series shrinkage module of the basic model is trained in the cloud, it uses Pearson correlation coefficient to screen time-series features related to COD and ammonia nitrogen load, retaining features with a correlation coefficient greater than or equal to 0.3 and eliminating redundant features to achieve feature dimension compression; the lightweight LSTM memory layer of the basic model reduces the number of model parameters and computational power consumption by simplifying the gated fully connected dimension of the standard LSTM.

[0014] As a further solution, the method of performing feedforward pre-regulation based on the predicted wastewater treatment load specifically includes adjusting the aeration frequency of the aeration device and the dosage of the chemical dosing device; wherein... Preset concentration thresholds for three levels: normal load, mild shock load, and severe shock load; When the predicted load reaches the threshold of severe shock load, the aeration frequency should be increased by 20%–35% and the continuous dosing mode should be turned on 4 hours in advance to increase the dosage of the agent. When the predicted load reaches the threshold of mild shock load, increase the aeration frequency by 10%–15% and increase the dosage of the chemical 2 hours in advance. When the predicted load is within the normal load range, the aeration frequency and the dosage of chemicals will be reduced back to the energy-saving benchmark value.

[0015] As a further solution, the feedback correction is combined with real-time effluent water quality parameters, including: collecting effluent COD and ammonia nitrogen data every 5 minutes, and when the deviation between the actual water quality and the predicted and regulated water quality exceeds a preset threshold, slightly adjusting the aeration frequency and the dosage of the reagent.

[0016] As a further solution, the rural-specific external fine-tuning factor data includes at least holiday time series labels, meteorological and rainfall data, village resident population base, and agritainment business data. The preprocessing of the rural-specific exogenous fine-tuning factor data includes: Holiday labels are digitized as 1, and weekday labels are digitized as 0; agritainment business hours are digitized as 1, and non-business hours are digitized as 0; rainfall status is divided into 0-3 levels according to rainfall level and coded; and the internal working condition data and all rural-specific external fine-tuning factor data have undergone anomaly removal, missing interpolation completion and normalization processing.

[0017] As a further solution, the predicted sewage treatment load for the next 2 hours and 4 hours is a fixed dual-time domain advance prediction, which is used to prepare in advance for the regular morning and evening sewage discharge peaks in rural areas and the severe impact loads caused by holidays and rainstorms.

[0018] On the other hand, the present invention also provides a multi-factor coupled intelligent control system for rural sewage load feedforward prediction, for executing a multi-factor coupled intelligent control method for rural sewage load feedforward prediction as described in any of the preceding claims, comprising: The cloud-based annual data collection and model pre-training module is used to collect the site's annual historical time-series data and static external information, train the lightweight time-series shrinking adaptive LSTM base model offline, and distribute the base model. The edge multi-source data acquisition module is used to collect internal working condition data and rural-specific external fine-tuning factor data in real time. The rural-specific exogenous factor encoding and fusion module is used to digitally encode and feature-fuse the rural-specific exogenous fine-tuning factor data to generate a multi-source fusion time-series feature matrix. The lightweight model incremental fine-tuning and prediction module is used to load the base model, freeze the network parameters of the time-series shrinking module and the lightweight LSTM memory layer, enable only the adaptive correction layer and use samples from the past 7 days to perform incremental self-learning on the adaptive correction layer, and use the updated model to predict the wastewater treatment load for the next 2 hours and 4 hours. The load level determination module is used to determine the current load level based on the predicted wastewater treatment load and the preset concentration threshold. The feedforward pre-regulation execution module is used to adjust the aeration frequency and the dosage of chemicals in advance according to the load level; The feedback correction closed-loop module is used to make minor corrections to the feedforward control amount based on real-time effluent water quality parameters. The local offline caching control module is used to cache running data when the network is interrupted, and to ensure that the prediction and control functions can run independently offline.

[0019] Compared with related technologies, the intelligent control method and system for feedforward prediction of rural sewage load provided by this invention has the following advantages: 1. This invention pioneers a 2–4 ​​hour dual-time-domain advanced prediction and hierarchical feedforward pre-control mechanism. Based on the annual long-term sewage discharge patterns obtained through cloud pre-training, it integrates real-time correction of four-dimensional rural-specific exogenous factors, including holidays, weather and rainfall, village population, and agritainment visitor flow, to predict peak influent loads in advance. This allows for pre-enhancement of aeration frequency and reagent dosage before high-concentration sewage enters the biochemical reaction unit, fundamentally eliminating the several-hour lag in response and adaptation of traditional feedforward control equipment to the biochemical microbial community. Application results show that during periods of shock load, the compliance rate of effluent COD and ammonia nitrogen increased from 72% to 99.2%, and the exceedance rate of effluent under shock conditions decreased by more than 90%.

[0020] 2. This invention pioneers a layered collaborative architecture where the cloud handles pre-training of the entire year's historical data, while the edge only handles incremental fine-tuning and real-time inference with small samples from the past 7 days. This offloads the computationally most demanding task of learning the entire dataset to the cloud, requiring only rolling cache of samples from the past 7 days at the edge gateway, eliminating the need for large-capacity local storage. Actual testing shows that the edge gateway's computing power utilization is less than 15%, allowing stable operation even with low-end NB-IoT embedded gateways, reducing hardware procurement costs by 33%. Furthermore, the entire process of prediction, incremental learning, and device control is permanently deployed locally on the edge gateway, ensuring uninterrupted functionality during network outages and automatic data resumption upon network recovery. This completely solves the control failure problem in rural areas with weak networks and intermittent network outages, making it fully adaptable to unattended scenarios without dedicated maintenance personnel.

[0021] 3. This invention uses long-term periodic emission patterns learned from year-round historical data in the cloud as the main model base. At the edge, four-dimensional rural-specific exogenous factors—holidays, weather and rainfall, village resident population, and agritourism visitor flow—are superimposed in real time as fine-tuning corrections. This constructs a hierarchical modeling logic of "long-term big data patterns as the main body and short-term scene disturbances as corrections," accurately depicting multiple load fluctuation patterns in rural areas, including morning and evening sewage discharge, holiday return trips, catering wastewater discharge, and rainwater runoff. This fills the technical gap of existing general urban models being unsuitable for rural pulse-like sewage discharge scenarios. Simultaneously, the edge adaptive correction layer incorporates a small-sample incremental self-learning mechanism from the past 7 days, automatically iterating and fine-tuning every 24 hours. This adapts to different village population sizes, sewage discharge habits, business types, and seasonal water temperature changes, eliminating the need for technicians to debug each site individually.

[0022] 4. This invention establishes a three-level load classification and control model. For severe impact loads, the aeration frequency is increased by 20%–35% and continuous dosing is activated 4 hours in advance. For mild impact loads, the aeration frequency is increased by 10%–15% 2 hours in advance. Normal loads automatically return to the energy-saving benchmark value. Combined with minute adjustments based on effluent quality feedback every 5 minutes, a dual closed-loop control system of feedforward coarse adjustment and feedback fine adjustment is formed. Application examples show that blower aeration energy consumption is reduced by 41%, and chemical consumption is reduced by 32%. While ensuring stable effluent compliance, significant energy-saving and consumption-reducing benefits are achieved, greatly reducing the operation and maintenance costs of rural wastewater treatment plants. Attached Figure Description

[0023] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0024] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart of the overall architecture of cloud + edge layered collaboration in this invention; Figure 2 This is a schematic diagram of the hierarchical structure of the lightweight temporal shrinkage adaptive LSTM model of the present invention; Figure 3 This is a flowchart of the multi-factor data acquisition and feature fusion process of the present invention; Figure 4 This is a logic block diagram of the three-level load feedforward pre-regulation strategy of the present invention. The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0027] Please see Figure 1 This embodiment provides a multi-factor coupled intelligent control method for feedforward prediction of rural sewage load, which is used for the automated control of rural decentralized sewage treatment plants. The method is characterized by including a cloud pre-training step and an edge local operation step. The cloud-based pre-training steps include: Obtain historical time-series data and static external source information of rural decentralized sewage treatment plants over the past year; The historical time-series data and the static external information of the site are preprocessed, and a lightweight time-series shrinkage adaptive LSTM basic model is trained offline based on the preprocessed historical time-series data to obtain a basic model containing the long-term annual sewage discharge pattern. The basic model is then distributed to the edge gateway of the corresponding site. The lightweight time-series shrinkage adaptive LSTM basic model includes a time-series shrinkage module, a lightweight LSTM memory layer, and an adaptive correction layer. The edge local operation steps include: Real-time collection of internal working condition data, and simultaneous acquisition of rural-specific external fine-tuning factor data; The intrinsic working condition data and the rural-specific extrinsic fine-tuning factor data are preprocessed to construct a multi-source fusion time-series feature matrix; The edge gateway loads the base model, freezes the network parameters of the time-series shrinking module and the lightweight LSTM memory layer, enables only the adaptive correction layer, and uses locally stored historical running samples to perform incremental self-learning on the adaptive correction layer to obtain an updated model adapted to local sewage discharge conditions. The multi-source fusion time-series feature matrix is ​​input into the update model to predict the wastewater treatment load for the next 2 and 4 hours. Based on the predicted wastewater treatment load, feedforward pre-control is implemented in advance, and feedback correction is performed in conjunction with real-time effluent water quality parameters, forming a dual closed-loop control of feedforward prediction and feedback correction.

[0028] The present invention will now be described in detail with reference to various embodiments: Example 1 This embodiment provides a multi-factor coupled intelligent control method for feedforward prediction of rural sewage load, which combines cloud pre-training and edge incremental self-learning. The method is executed in a hierarchical and collaborative manner through cloud and edge gateways, and consists of two main parts: cloud pre-training steps and edge local operation steps.

[0029] Figure 1 This is a flowchart illustrating the overall cloud-edge layered collaborative architecture of the present invention, clearly showing the data flow and functional division between the cloud training layer and the edge local execution layer. The following is combined with... Figure 1 Each step is explained in detail.

[0030] I. Cloud-based pre-training steps This step corresponds to Figure 1The solution involves the cloud-based data center and the S0 stage. In rural decentralized sewage treatment scenarios, existing technologies typically deploy all historical data storage and model training locally on the edge gateway, or rely entirely on the cloud for real-time inference. The former requires the edge gateway to have large-capacity storage and high-performance computing power, resulting in high hardware costs that low-end embedded gateways cannot handle. The latter is highly dependent on real-time network connectivity. In rural environments with weak 4G / NB-IoT signals and intermittent network outages, cloud transmission latency is high, data packet loss is frequent, and on-site equipment completely loses its control capabilities during network outages. To resolve the contradiction between the above-mentioned computing power deployment and network conditions, this embodiment adopts a layered architecture where the cloud undertakes full-year data pre-training, and the edge is only responsible for small-sample incremental fine-tuning and real-time inference. The task of learning all historical data, which consumes the most computing power, is offloaded to a high-performance server in the cloud, significantly reducing the storage and computing power threshold at the edge.

[0031] The specific process is as follows: First, such as Figure 1 As shown in the cloud-based data center section, the cloud platform centrally accesses the historical databases of all decentralized rural sewage treatment plants within the jurisdiction, performing year-round historical data collection to obtain historical time-series data for each plant over the past year. This historical time-series data specifically includes the daily sewage treatment volume, daily influent flow rate, and water quality time-series data such as influent COD concentration and ammonia nitrogen concentration for each plant over the past year. Simultaneously, the cloud platform synchronously acquires and stores static external source information for each plant. This static external source information includes at least the village's permanent resident population, the region's average annual rainfall characteristics, and the basic operating scale of agritourism businesses within the jurisdiction.

[0032] Then, as Figure 1 As shown in the medium-batch preprocessing steps, the cloud performs preprocessing on the aforementioned historical time-series data. The preprocessing process includes: using the 3σ criterion to remove abnormal sensor jump data; using temporal proximity interpolation to fill in short-term missing data; using temporal smoothing to reduce noise caused by equipment maintenance and short-term abnormal interference; and performing min-max normalization on the entire time-series data, mapping it to the [0,1] interval. For static external information from the site, the cloud organizes and digitally stores it for subsequent distribution.

[0033] After preprocessing is completed, such as Figure 1 As shown in the steps of full pre-training of the lightweight LSTM, the lightweight time-series shrinking adaptive LSTM base model (hereinafter referred to as the base model) is trained offline in the cloud based on the pre-processed complete annual historical time-series data.

[0034] Figure 2 This is a schematic diagram of the hierarchical structure of the lightweight temporal shrinkage adaptive LSTM basic model of this invention. (Refer to...) Figure 2The basic model consists of three core modules connected in sequence: a temporal shrinkage module, a lightweight LSTM memory layer, and an adaptive correction layer. The input multi-source features are first compressed and filtered by the temporal shrinkage module, then the temporal dependency patterns are extracted by the lightweight LSTM memory layer, and finally the prediction results (including wastewater treatment volume, influent COD peak value, and ammonia nitrogen peak value) are output through the adaptive correction layer.

[0035] In existing technologies, traditional LSTM models used for wastewater prediction typically use all time-series features directly as input without feature selection, and the model structure has not undergone targeted lightweight modifications. This leads to two problems: first, many time-series features in rural wastewater data have weak correlations with the prediction targets (COD, ammonia nitrogen load), and redundant features not only increase computational overhead but may also introduce noise, reducing prediction accuracy; second, standard LSTM models have a large number of parameters and high computational consumption, making them unable to run smoothly on low-end embedded gateways. To address these issues, this embodiment performs two structural modifications to the basic model.

[0036] in, Figure 2 The time-series shrinkage module in the cloud training process calculates the Pearson correlation coefficient between each time-series feature and future COD and ammonia nitrogen loads. It then filters out strongly correlated features with an absolute correlation coefficient greater than or equal to 0.3, eliminating weakly correlated or redundant time-series features. This compresses the feature dimension by 30%–50%, reducing the input complexity of subsequent model calculations. The aim of this design is to filter out noisy features that contribute little to the prediction at the source, allowing the model to focus on learning the core factors that truly affect wastewater treatment loads, thereby improving prediction accuracy and generalization ability.

[0037] Figure 2 The lightweight LSTM memory layer in the model significantly simplifies the fully connected dimension in the gating structure of the standard LSTM network, reducing the number of hidden layer nodes and connection parameters. After this lightweight transformation, the overall model's parameter count and computational consumption are reduced by more than 60% compared to the standard LSTM, enabling it to run smoothly on low-computing-power embedded edge gateways. This design allows deep learning models that originally required high-performance servers to be compressed and deployed in cost-sensitive, computationally limited rural edge gateways, creating conditions for large-scale batch deployment.

[0038] The cloud-based system fully trains the base model using year-round data, enabling it to learn the annual seasonal patterns, monthly trends, and daily morning and evening peak characteristics of wastewater discharge from the monitoring station. After training convergence, the model weight file is exported, such as... Figure 1As shown in the steps of distributing model weights, the basic model, along with the previously compiled static external source information of the sites, is distributed to the local persistent storage of the corresponding site's edge gateway. By learning long-term patterns throughout the year in the cloud, the model gains the ability to accurately characterize the baseline trend of site pollution discharge, which provides a solid prior knowledge base for subsequent accurate short-term predictions at the edge.

[0039] In addition, the cloud platform has an annual iteration and update mechanism. Every 12 months, the cloud automatically aggregates the newly added complete annual historical time-series data from each site, re-executes the aforementioned preprocessing and full offline training processes, generates an updated base model, and remotely distributes it to the edge gateway at an opportune time to replace the old model, continuously maintaining an accurate portrayal of the long-term sewage discharge patterns of the sites. Factors such as the permanent population of rural villages, the scale of agritainment businesses, and villagers' water usage habits will slowly evolve over time. The annual iteration mechanism ensures that the base model always reflects the latest long-term trends, avoiding the degradation of prediction accuracy caused by model aging.

[0040] II. Edge Local Operation Steps Return to reference Figure 1 After receiving the basic model from the cloud, the edge gateway is responsible for completing all subsequent tasks, including data collection, incremental learning, load forecasting, and real-time control. Figure 1 The edge gateway section of the project illustrates the entire process of multi-source data acquisition and preprocessing, external factor fusion, incremental adaptive correction, 2h / 4h load forecasting, load classification determination, feedforward pre-regulation and feedback correction, and finally, equipment execution. In existing technologies, rural wastewater treatment plant control generally employs feedforward closed-loop logic, which involves collecting real-time water quality parameters through online sensors and passively adjusting the equipment when water quality exceeds standards. This approach inherently suffers from lag: when sensors detect abnormal water quality, high-concentration wastewater has already entered the biochemical reaction unit, and equipment response and bacterial adaptation take several hours, inevitably leading to short-term effluent exceeding standards. Furthermore, rural wastewater discharge exhibits typical "time-period pulse-like and festival-burst" characteristics. During peak morning and evening washing hours, holiday return trips, concentrated operation of agritourism businesses, and rainfall, the impact load intensity is high and concentrated in time, making feedforward control completely ineffective. To address the aforementioned control lag issue, this embodiment establishes a feedforward prediction mechanism of "early prediction + pre-control" at the edge side, which completes the process parameter enhancement before high-concentration wastewater enters the biochemical unit, thus preventing excessive shock effluent from the source.

[0041] All operational logic on the edge gateway supports independent offline operation, without relying on real-time cloud computing power. The local edge operation process can be divided into five sub-stages, S1 to S5, each sub-stage and... Figure 1 The modules within the edge gateway are in a corresponding relationship.

[0042] (S1) Multi-source data acquisition and preprocessing Existing predictive modeling schemes mostly rely on single-type time-series data of water quality and flow collected by instruments within the station, resulting in a limited information source. However, fluctuations in rural sewage discharge cannot be fully characterized by station parameters alone; they are substantially related to external social and environmental factors such as the village's resident population, holiday periods, weather and rainfall, and customer flow from rural tourism and catering businesses. Current technologies fail to incorporate these village-specific disturbances during modeling, leading to a significant drop in prediction accuracy when data volume is low or sewage discharge patterns undergo temporary changes, resulting in insufficient adaptability and generalization capabilities. To address this issue, this embodiment specifically introduces village-specific exogenous fine-tuning factors at the edge, which, together with the station data, constitute a multi-source fusion input, improving prediction accuracy and scenario adaptability from the data source.

[0043] Figure 3 This is a flowchart illustrating the multi-factor data acquisition and feature fusion process of this invention. (Refer to...) Figure 3 The data sources are divided into three categories: annual historical baseline time-series data from the cloud, real-time internal operating condition data from the edge, and external fine-tuning factor data from the edge (specifically including four categories: holiday tags, weather and rainfall, village population, and agritainment visitor flow). After preprocessing and fusion, the three types of data output a multi-source fused time-series feature matrix for use in subsequent prediction models.

[0044] The following combination Figure 3 The data sources and integration process are explained in detail.

[0045] The edge gateway no longer stores a complete historical time-series database locally for an entire year; instead, it only caches locally run samples from the past 7 days for incremental self-learning. The edge gateway synchronously collects the following multi-source data at 5-minute intervals: (1) Internal operating condition data: including influent instantaneous flow rate, influent COD concentration, influent ammonia nitrogen concentration, biological tank DO concentration, MLSS sludge concentration, current operating frequency of blower, start and stop status of dosing pump, etc.

[0046] (2) Rural-specific external fine-tuning factor data: Specifically, it includes four categories: holiday time series labels for the current day and the next 4 hours (weekdays / holidays), real-time rainfall meteorological data, static base of village permanent residents, and real-time customer flow and business hours of agritainment.

[0047] After collecting the above data, the edge gateway performs standardized preprocessing on all data. For internal operating condition data, the 3σ criterion is used to remove abnormal jump values, and time-series proximity interpolation is used to fill in missing values ​​caused by short-term network outages or sensor failures. Min-max normalization is then performed to the [0,1] interval. For rural-specific external fine-tuning factor data, the following digitization is performed: holiday labels are digitized as 1, weekday labels as 0; agritainment business hours are digitized as 1, non-business hours as 0; rainfall status is divided into 0–3 levels according to rainfall intensity and coded accordingly. The coded external factor data also undergoes min-max normalization.

[0048] After the above preprocessing, as Figure 3 The final output shows that the long-term trend information contained in the annual historical benchmark patterns in the cloud, the internal working condition data collected in real time at the edge, and the encoded rural-specific external fine-tuning factor data are combined to construct a multi-source fusion time-series feature matrix, which serves as the input for subsequent incremental learning and prediction. This multi-source fusion method takes long-term big data patterns as the main body and short-term scene disturbances as corrections, enabling the model to maintain a stable portrayal of annual, monthly, and intraday cyclical patterns while responding sensitively to temporary external disturbances, thus balancing long-term trend accuracy with short-term shock response sensitivity.

[0049] (S2) Edge Increment Adaptive Correction Rural wastewater treatment plants are widely distributed, and their operating conditions vary significantly across different villages due to differences in resident population size, villagers' water usage habits, agritourism business models, and seasonal water temperature variations. Relying solely on a general, cloud-based model for prediction cannot adapt to the individual differences of each plant, leading to a gradual accumulation of prediction errors. Traditional methods require specialized technicians to fine-tune parameters and retrain the model for each plant, resulting in extremely high maintenance costs. To address this issue, this embodiment incorporates a dedicated adaptive correction layer within the basic model and designs a fully automated incremental self-learning mechanism.

[0050] Refer again Figure 2 The edge gateway loads the lightweight temporal shrinking adaptive LSTM base model delivered from the cloud. To minimize the computational overhead at the edge, the edge gateway freezes all network parameters of the temporal shrinking module and the lightweight LSTM memory layer in the base model, only enabling... Figure 2 The adaptive correction layer in the core performs subsequent incremental fine-tuning of its shallow weights. By freezing the backbone network, incremental learning only requires updating a very small number of shallow parameters, significantly reducing computational power requirements and enabling stable operation even for low-end embedded gateways.

[0051] Figure 2The adaptive correction layer incorporates an incremental self-learning mechanism based on small samples from the past 7 days. Every 24 hours, the edge gateway automatically reads the historical operating samples from the past 7 days cached locally (these samples are data collected during the qualifying operating period) and performs a lightweight iterative fine-tuning. The fine-tuning process only updates the weights of the adaptive correction layer slightly, correcting prediction biases caused by factors such as short-term holidays, heavy rain, temporary fluctuations in passenger flow, seasonal water temperature changes, or adjustments in rural business formats. This allows the model to quickly adapt to the real-time sewage discharge characteristics of the local site while maintaining its long-term trend memory capability. This incremental self-learning process is completed entirely locally on the edge gateway, unaffected by network connectivity and requiring no manual intervention, achieving "one model per site" adaptive adaptation and significantly reducing maintenance complexity.

[0052] (S3) Dual-time domain load advance prediction and level determination Rural sewage shock loads exhibit significant temporal differences: peak discharge times during the morning and evening, when villagers concentrate on washing and laundry, have a relatively short lag time of approximately 2 hours from generation to entry into the pipe network and then to the treatment plant; however, severe shock loads caused by holiday travel, concentrated dining at agritourism establishments, and heavy rainstorms result in longer collection and transmission times in the pipe network, leading to greater impact intensity and requiring longer advance response time. Existing technologies lack the ability to differentiate between these two timescale shock loads and provide targeted responses. To address this issue, this embodiment establishes a fixed 2-hour and 4-hour dual-time-domain advance prediction mechanism.

[0053] Once adaptive calibration is complete and the model is in its latest state, the edge gateway inputs the real-time updated multi-source fusion time-series feature matrix into the updated model. The model, based on the annual long-term cycle basis learned in the cloud, fuses real-time internal operating conditions at the edge and rural-specific external fine-tuning factor information, outputting prediction results for two key time domains: the next 2 hours and the next 4 hours. These prediction results include, but are not limited to: predicted wastewater treatment volume, peak influent COD, peak influent ammonia nitrogen, and maximum instantaneous flow rate.

[0054] Based on this, the edge gateway determines the level of the predicted load according to the preset three-level concentration thresholds. The specific settings of the three-level concentration thresholds are as follows: normal load corresponds to influent COD concentration ≤150mg / L and ammonia nitrogen concentration ≤15mg / L; mild shock load corresponds to 150mg / L < COD concentration ≤250mg / L, or 15mg / L < ammonia nitrogen concentration ≤25mg / L; severe shock load corresponds to COD concentration >250mg / L, or ammonia nitrogen concentration >25mg / L. Combining the predicted wastewater treatment volume and the above concentration thresholds, the edge gateway marks the corresponding load level for the influent load in the current and future periods, providing a precise basis for subsequent graded pre-control. This graded method maps the continuously changing load intensity to three discrete control levels, enabling subsequent feedforward pre-control to match the corresponding intensity of process parameter reinforcement measures according to the severity of the load, avoiding the crudeness of "one-size-fits-all" control.

[0055] (S4) Hierarchical feedforward pre-regulation and feedback correction dual closed-loop control Existing technologies mostly employ fixed-level or single-threshold control methods, failing to match process parameters in advance based on actual load changes. During low-load periods, equipment may still operate at higher parameters, resulting in ineffective consumption of electricity and chemicals; during periods of high-impact loads, the equipment's control capabilities are insufficient to cope with the risk of exceeding limits. While single feedforward control solves the lag problem, it is limited by the accuracy of the prediction model itself, and prediction bias may still exist, leading to a mismatch between the control amount and actual demand. To address these issues, this embodiment establishes a dual-closed-loop adaptive control logic that prioritizes feedforward prediction and supplements it with feedback correction, balancing rapid response and precise control to achieve a dynamic balance between compliant operation and energy conservation.

[0056] Figure 4 This is a logic block diagram of the three-level load feedforward pre-regulation strategy of the present invention. The following is in conjunction with... Figure 4 A detailed explanation will be provided.

[0057] Reference Figure 4 The system input is the load level predicted in the S3 stage, and three differentiated control paths are executed according to the different load levels: (1) such as Figure 4 As shown in the left-hand branch, when the predicted load reaches the level of a severe impact load (such as major holidays, rainstorms, concentrated operation of agritainment businesses, peak travel seasons, etc.), the aeration frequency of the aeration blowers is increased 4 hours in advance, with an increase of 20%–35% (this parameter can be configured remotely) to increase the dissolved oxygen reserve in the biological treatment tank; at the same time, the micro-continuous dosing mode is activated to increase the base concentration of nitrogen and phosphorus removal agents, so that the biological treatment system has sufficient treatment capacity before high-concentration wastewater arrives. The execution results are as follows: Figure 4 The aeration volume and the dosage of the pesticide shown in the figure increased significantly.

[0058] (2) For example Figure 4 As shown in the middle branch, when the predicted load reaches the level of a mild shock load (such as the fixed peak sewage discharge generated by villagers washing and doing laundry in the morning and evening), the aeration frequency is slightly increased 2 hours in advance, with an increase of 10%–15% (parameters are configurable), and the pulse dosage of the reagent is finely adjusted to balance short-term water quality fluctuations. The execution results are as follows: Figure 4 The aeration rate and dosage shown are slightly increased.

[0059] (3) such as Figure 4 As shown in the right-hand branch, when the predicted load is at the normal load level, the aeration frequency and chemical dosage are automatically reduced back to the energy-saving benchmark value, and redundant intermittent dosing is shut down to avoid over-aeration and over-dosing during low-load periods, thereby reducing ineffective energy and chemical consumption. The execution result is as follows: Figure 4 The aeration rate and dosage are kept to a minimum, as shown in the diagram.

[0060] like Figure 4 As shown at the bottom, after dynamic execution, the results of the three types of regulation are all connected to the effluent water quality feedback correction closed loop. Specifically, the edge gateway collects the effluent COD concentration and ammonia nitrogen concentration every 5 minutes, comparing the actual effluent water quality with the expected water quality after prediction and regulation. If the deviation exceeds the preset threshold, the current aeration frequency and reagent dosage are slightly adjusted to eliminate the impact of prediction deviation or external unmodeled interference, achieving precise closed-loop control. Through the dual closed-loop collaboration of feedforward coarse adjustment for rapid response to load changes and feedback fine adjustment to eliminate residual deviations, the system can not only significantly anticipate shock loads but also ensure accurate and stable long-term operation.

[0061] (S5) Cloud-edge data synchronization and offline independent operation mechanism The widespread deployment of 4G and NB-IoT networks in rural areas suffers from weak signal coverage, low bandwidth, and intermittent network outages. Pure cloud-based solutions completely lose their predictive and control capabilities during network interruptions, making real-time on-site management impossible. To address this issue, this embodiment embeds all core logic for prediction, incremental learning, and device control locally on the edge gateway, ensuring uninterrupted functionality even under network outage conditions.

[0062] Return to reference Figure 1 , Figure 1 The bottom clearly shows the data flow path in both online and offline states. Throughout operation, the edge gateway possesses comprehensive online synchronization and offline independent operation capabilities.

[0063] like Figure 1As shown by the dotted arrow indicating the data transmission from the edge gateway to the cloud for archiving during the China Internet Network period, when the network connection is normal, the edge gateway regularly uploads the incremental samples of the past 7 days, equipment operation logs, and water quality data stored locally to the cloud for archiving daily, providing a data foundation for global retraining in the cloud the following year. Simultaneously, the cloud can remotely distribute updated basic models or revised control threshold parameters.

[0064] When network outages occur (such as weak 4G / NB-IoT signals or intermittent network disconnections in rural areas), the edge gateway relies on a local caching mechanism to temporarily store nearly seven days' worth of complete operational data locally. All control logic, including incremental self-learning, load forecasting, tiered pre-regulation, and device driving, continues to operate independently without any functional interruption. Upon network recovery, the edge gateway automatically resumes interrupted data transmission, synchronizing all cached data from the interruption period to the cloud. This offline independent operation capability completely solves the industry pain point of control failure caused by unstable rural networks, enabling the system to truly adapt to unattended scenarios at the village and township levels with poor network conditions and no dedicated maintenance personnel.

[0065] Example 2 This embodiment provides a multi-factor coupled edge intelligent control system for rural sewage treatment, combining cloud-based pre-training and edge incremental self-learning. This system executes all the steps of the method described in Embodiment 1. The overall system architecture is as follows: Figure 1 As shown, the structure mainly consists of a cloud component and an edge gateway component.

[0066] The cloud-based component includes a cloud-based annual data collection and model pre-training module. This module collects the year-round historical time-series data and static external information of each site, performs batch preprocessing on the historical time-series data, trains a lightweight time-series shrinking adaptive LSTM base model offline, and then distributes the trained base model along with the site's static external information to the corresponding edge gateway. This module corresponds to... Figure 1 The full functionality of the cloud-based data center.

[0067] The edge gateway is internally divided into the following modules based on function, and each module corresponds to... Figure 1 The functional units shown inside the field edge gateway are as follows: The edge multi-source data acquisition module is used to collect real-time internal operating condition data and village-specific external fine-tuning factor data in a 5-minute cycle. Internal operating condition data includes influent instantaneous flow rate, influent COD concentration, influent ammonia nitrogen concentration, biological treatment tank DO concentration, MLSS sludge concentration, blower operating frequency, and dosing pump start / stop status. Village-specific external fine-tuning factor data includes at least holiday time series labels, meteorological rainfall data, village resident population base, and agritourism visitor flow data. This module corresponds to... Figure 1 Multi-source data acquisition and preprocessing functions.

[0068] The rural-specific exogenous fine-tuning factor encoding and fusion module is used to digitally encode and fuse the aforementioned rural-specific exogenous fine-tuning factor data, generating a multi-source fused time-series feature matrix. This module corresponds to... Figure 1 The exogenous factor fusion function in the data fusion process is as follows: Figure 3 As shown.

[0069] The lightweight model incremental fine-tuning and prediction module loads the base model distributed from the cloud, freezes the network parameters of the time-series shrinking module and the lightweight LSTM memory layer, and only enables the adaptive calibration layer. It uses locally cached historical running samples from the past 7 days to perform daily incremental self-learning on the adaptive calibration layer, obtaining an updated model adapted to local sewage discharge conditions. The real-time multi-source fusion time-series feature matrix is ​​then input into this updated model to predict the sewage treatment load for the next 2 and 4 hours. The prediction results include at least the sewage treatment volume, peak influent COD, peak ammonia nitrogen, and maximum instantaneous flow rate. This module corresponds to... Figure 1 The incremental adaptive correction and 2h / 4h load forecasting functions in the model have the following internal structure: Figure 2 As shown.

[0070] The load level determination module is used to determine the load level for the current and future periods based on the predicted wastewater treatment load and preset concentration thresholds for three levels: normal load, mild shock load, and severe shock load. This module corresponds to... Figure 1 The load classification determination function in the system.

[0071] The feedforward pre-regulation module is used to adjust the aeration frequency and pesticide dosage in advance according to the load level. Specifically, for severe impact loads, the aeration frequency is increased by 20%–35% and continuous dosing is activated 4 hours in advance; for mild impact loads, the aeration frequency is increased by 10%–15% and the pesticide dosage is increased 2 hours in advance; and for normal loads, the aeration frequency is reduced to the energy-saving baseline value. This module corresponds to… Figure 1 The feedforward pre-regulation function in the middle, its regulation logic is as follows: Figure 4 As shown.

[0072] The feedback correction closed-loop module collects effluent COD and ammonia nitrogen data every 5 minutes. When the deviation between the actual water quality and the predicted and adjusted values ​​exceeds a preset threshold, it makes minor adjustments to the aeration frequency and reagent dosage, forming a dual closed-loop control of feedforward prediction and feedback correction. This module corresponds to... Figure 1 The feedback correction function in the middle has a closed-loop logic as follows: Figure 4 As shown at the bottom.

[0073] The local offline caching control module automatically caches complete operational data for the past 7 days when the network is interrupted, ensuring that incremental learning, load forecasting, feedforward pre-regulation, and feedback correction functions all run independently and continuously on the edge gateway. Once the network is restored, the cached data is re-uploaded to the cloud. This module corresponds to... Figure 1 Offline caching function that runs independently in the middle edge gateway when the network is offline.

[0074] The above modules work together, and the method described in the complete embodiment 1 can achieve the same technical effect as that in embodiment 1, so it will not be repeated here.

[0075] Example 3 This embodiment takes a township-level decentralized A² / O integrated rural sewage treatment station as the implementation object. The station is designed to treat 50 m³ / d, with no dedicated operation and maintenance personnel on duty. Only a low-end edge smart gateway and conventional water quality sensors are deployed on site, and the communication environment is a low-bandwidth NB-IoT network.

[0076] The cloud platform first collects the daily wastewater treatment volume and influent water quality time-series data for the entire past year for this site. After unified preprocessing, it performs full pre-training of a lightweight time-series shrinking adaptive LSTM basic model to fully learn the daily, monthly, and seasonal wastewater discharge baseline patterns of the site. The site's static external information includes: a permanent resident population of 200 people and the basic operating scale of 3 agritainment businesses within the village. After cloud training is completed, the model weights and the aforementioned static external information are distributed to the on-site edge gateway.

[0077] The edge gateway is configured with a 5-minute cycle for data collection, internally collecting parameters such as influent COD, ammonia nitrogen, instantaneous flow rate, DO, and fan frequency. It also receives real-time data on local weather and rainfall, holiday tags, and agritourism operating hours and visitor flow. The gateway only stores samples from the past 7 days on a rolling basis. After loading the basic model, the backbone network is frozen, and adaptive correction is performed every 24 hours based on samples that meet the standards from the past 7 days. Actual testing showed that the edge gateway's computing power utilization rate was less than 15%, fully adaptable to low-end embedded hardware. This verifies the significant effect of the layered architecture—where the cloud handles full pre-training and the edge only handles small-sample incremental fine-tuning—in lowering the hardware barrier for edge applications.

[0078] In actual operation, during weekday morning and evening hours of 7:00–9:00 and 19:00–21:00, the model predicts mild shock loads and increases the aeration frequency by 12% two hours in advance, slightly increasing the pulse dosage of chemicals to smoothly cope with the concentrated sewage discharge from villagers' washing and laundry. On weekends and statutory holidays, the system automatically identifies the external source label of holidays and predicts severe shock loads, increasing the aeration frequency by 30% four hours in advance and starting the continuous micro-dosing mode to adapt to the surge in returning migrant workers and the concentrated sewage discharge from agritainment restaurants. During rainy weather, after integrating real-time rainfall external source factors, the model predicts the shock load brought by pipeline scouring, automatically upgrades the shock level, and strengthens process parameters in advance to avoid exceeding the effluent standards. The above-mentioned control behavior in different scenarios verifies the effectiveness of the multi-factor coupled prediction and hierarchical feedforward pre-control mechanism under various typical rural shock conditions.

[0079] After 30 consecutive days of operation, the effluent COD and ammonia nitrogen compliance rates at this site increased from 72% to 99.2% during periods of high load, validating the ability of the feedforward prediction mechanism to fundamentally solve the lag problem of traditional feedforward control. Aeration energy consumption by the blower decreased by 41%, and reagent consumption decreased by 32%, validating the significant energy-saving and consumption-reducing effects of the three-level hierarchical control in ensuring compliance under high load and reducing energy consumption under low load. During a 7-day continuous network outage test, the incremental learning, load prediction, and hierarchical control logic of the edge gateway remained uninterrupted and functioned flawlessly, validating the reliability of the offline independent operation mechanism. The edge gateway does not require large-capacity local storage, reducing hardware procurement costs by 33% compared to a fully local training solution, validating the deployment cost advantages of the cloud-edge layered architecture. The annual model pre-training and remote distribution for dozens of village and town sites within the jurisdiction were completed in a unified batch on the cloud, eliminating the need for on-site debugging at each site, significantly reducing maintenance workload and fully demonstrating the outstanding advantages of this method in rural weak network, unattended, and low-cost large-scale deployment scenarios.

[0080] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the content of this application specification and drawings, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A multi-factor coupled intelligent control method for feedforward prediction of rural sewage load, used for the automated control of rural decentralized sewage treatment plants, characterized in that, This includes cloud-based pre-training steps and edge-based local execution steps; The cloud-based pre-training steps include: Obtain historical time-series data and static external source information of rural decentralized sewage treatment plants over the past year; The historical time-series data and the static external information of the site are preprocessed, and a lightweight time-series shrinkage adaptive LSTM basic model is trained offline based on the preprocessed historical time-series data to obtain a basic model containing the long-term annual sewage discharge pattern. The basic model is then distributed to the edge gateway of the corresponding site. The lightweight time-series shrinkage adaptive LSTM basic model includes a time-series shrinkage module, a lightweight LSTM memory layer, and an adaptive correction layer. The edge local operation steps include: Real-time collection of internal working condition data, and simultaneous acquisition of rural-specific external fine-tuning factor data; The intrinsic working condition data and the rural-specific extrinsic fine-tuning factor data are preprocessed to construct a multi-source fusion time-series feature matrix; The edge gateway loads the base model, freezes the network parameters of the time-series shrinking module and the lightweight LSTM memory layer, enables only the adaptive correction layer, and uses locally stored historical running samples to perform incremental self-learning on the adaptive correction layer to obtain an updated model adapted to local sewage discharge conditions. The multi-source fusion time-series feature matrix is ​​input into the update model to predict the wastewater treatment load for the next 2 and 4 hours. Based on the predicted wastewater treatment load, feedforward pre-control is implemented in advance, and feedback correction is performed in conjunction with real-time effluent water quality parameters, forming a dual closed-loop control of feedforward prediction and feedback correction.

2. The intelligent control method for feedforward prediction of rural sewage load based on multi-factor coupling according to claim 1, characterized in that, The cloud-based pre-training step further includes: every 12 months, summarizing the newly added historical time-series data for the entire year at the site, re-executing the preprocessing and offline training to obtain an updated base model, and remotely distributing the updated base model to the edge gateway of the corresponding site.

3. The intelligent control method for feedforward prediction of rural sewage load based on multi-factor coupling according to claim 1, characterized in that, The edge gateway only saves local operating samples for the past 7 days on a rolling basis, and the incremental self-learning, the prediction of wastewater treatment load, and the feedforward pre-control are all performed locally on the edge gateway. When the network connection is normal, the edge gateway will synchronously upload local operating data and samples to the cloud for archiving; When the network is interrupted, the edge gateway independently completes prediction and control based on local cached data, and resumes the transmission of cached data after the network is restored.

4. The intelligent control method for feedforward prediction of rural sewage load based on multi-factor coupling according to claim 3, characterized in that, The adaptive correction layer of the basic model adopts an incremental self-learning mechanism based on small samples from the past 7 days. Every 24 hours, it reads the running samples from the past 7 days in the local cache for lightweight iterative fine-tuning, automatically adapting to the sewage discharge characteristics, seasonal water temperature and business changes of different villages.

5. The intelligent control method for feedforward prediction of rural sewage load based on multi-factor coupling according to claim 1, characterized in that, The wastewater treatment load includes at least the wastewater treatment volume, peak influent COD, peak ammonia nitrogen, and maximum instantaneous flow rate. When the time-series shrinkage module of the basic model is trained in the cloud, it uses Pearson correlation coefficient to screen time-series features related to COD and ammonia nitrogen load, retaining features with a correlation coefficient greater than or equal to 0.3 and eliminating redundant features to achieve feature dimension compression. The lightweight LSTM memory layer of the basic model reduces the number of model parameters and computational power consumption by simplifying the gated fully connected dimension of the standard LSTM.

6. The intelligent control method for feedforward prediction of rural sewage load based on multi-factor coupling according to claim 1, characterized in that, The aforementioned pre-control based on the predicted wastewater treatment load specifically includes adjusting the aeration frequency of the aeration device and the dosage of the chemical dosing device; wherein... Preset concentration thresholds for three levels: normal load, mild shock load, and severe shock load; When the predicted load reaches the threshold of severe shock load, the aeration frequency should be increased by 20%–35% and the continuous dosing mode should be turned on 4 hours in advance to increase the dosage of the agent. When the predicted load reaches the threshold of mild shock load, increase the aeration frequency by 10%–15% and increase the dosage of the chemical 2 hours in advance. When the predicted load is within the normal load range, the aeration frequency and the dosage of chemicals will be reduced back to the energy-saving benchmark value.

7. The intelligent control method for feedforward prediction of rural sewage load based on multi-factor coupling according to claim 6, characterized in that, The feedback correction based on real-time effluent water quality parameters includes: collecting effluent COD and ammonia nitrogen data every 5 minutes; and making minor adjustments to the aeration frequency and reagent dosage when the deviation between the actual water quality and the predicted and regulated water quality exceeds a preset threshold.

8. The intelligent control method for feedforward prediction of rural sewage load based on multi-factor coupling according to claim 1, characterized in that, The rural-specific external fine-tuning factor data includes at least holiday time series labels, meteorological and rainfall data, village resident population base, and agritainment business data. The preprocessing of the rural-specific exogenous fine-tuning factor data includes: Holiday labels are digitized as 1, and weekday labels are digitized as 0; agritainment business hours are digitized as 1, and non-business hours are digitized as 0; rainfall status is divided into 0-3 levels according to rainfall level and coded; and the internal working condition data and all rural-specific external fine-tuning factor data have undergone anomaly removal, missing interpolation completion and normalization processing.

9. A multi-factor coupled intelligent control method for feedforward prediction of rural sewage load according to any one of claims 6 and 8, characterized in that, The predicted wastewater treatment load for the next 2 hours and 4 hours is a fixed dual-time domain advance prediction, which is used to prepare in advance for the regular morning and evening sewage discharge peaks in rural areas and the severe impact loads caused by holidays and rainstorms.

10. A multi-factor coupled intelligent control system for rural sewage load feedforward prediction, used to execute the multi-factor coupled intelligent control method for rural sewage load feedforward prediction as described in any one of claims 1 to 9, characterized in that, include: The cloud-based annual data collection and model pre-training module is used to collect the site's annual historical time-series data and static external information, train the lightweight time-series shrinking adaptive LSTM base model offline, and distribute the base model. The edge multi-source data acquisition module is used to collect internal working condition data and rural-specific external fine-tuning factor data in real time. The rural-specific exogenous factor encoding and fusion module is used to digitally encode and feature-fuse the rural-specific exogenous fine-tuning factor data to generate a multi-source fusion time-series feature matrix. The lightweight model incremental fine-tuning and prediction module is used to load the base model, freeze the network parameters of the time-series shrinking module and the lightweight LSTM memory layer, enable only the adaptive correction layer and use samples from the past 7 days to perform incremental self-learning on the adaptive correction layer, and use the updated model to predict the wastewater treatment load for the next 2 hours and 4 hours. The load level determination module is used to determine the current load level based on the predicted wastewater treatment load and the preset concentration threshold. The feedforward pre-regulation execution module is used to adjust the aeration frequency and the dosage of chemicals in advance according to the load level; The feedback correction closed-loop module is used to make minor corrections to the feedforward control amount based on real-time effluent water quality parameters. The local offline caching control module is used to cache running data when the network is interrupted, and to ensure that the prediction and control functions can run independently offline.

Images