Assembly type aerobic granular sludge reactor operation parameter self-adaption method and system

Abstract

The invention relates to the technical field of sewage treatment, discloses a self-adaptive method and system for operating parameters of an assembled aerobic granular sludge reactor, and aims to solve the problems of unstable treatment efficiency, high energy consumption and substandard effluent due to the fact that existing operation management depends on presetting or empirical adjustment. The method comprises the steps of collecting and preprocessing multi-dimensional operation data in real time to construct a multi-modal feature set; predicting future key performance indexes through a deep learning model; determining optimal operation parameters (a total operation period, a sludge backflow proportion, aeration intensity and sludge-water mixing time) by utilizing a multi-objective optimization algorithm, and executing an instruction; and feeding back monitoring data to iteratively update the model. The system comprises a data acquisition module, a preprocessing module, a feature engineering module, a deep learning prediction module, a multi-objective optimization module, an instruction generation module and a feedback learning module. By adopting the technical scheme, the intelligent and automatic level of the reactor can be remarkably improved, the effluent is ensured to reach the standard, the energy consumption and the carbon source consumption are reduced, and the reactor has long-term stable operation and environment self-adaptive capability.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of sewage treatment, and particularly relates to a method and system for self-adapting operation parameters of an assembled aerobic granular sludge reactor. BACKGROUND

[0002] With the increasing strictness of water treatment technology on efficiency and environmental standards, the aerobic granular sludge technology is becoming an important solution to improve treatment effect, reduce land occupation and reduce sludge production in the field of municipal and industrial sewage treatment due to its high biomass, high settling performance and strong impact load resistance. The technology promotes the formation of dense granules by microbial flocculation, significantly enhances the settling performance and activity of sludge, and is considered as one of the core technologies of future sewage treatment.

[0003] Among them, the assembled aerobic granular sludge reactor, even if it integrates a reinforced biological combined system such as P pool, B pool and SU pool, still generally relies on preset standard cycles, empirical adjustments or control strategies based on simple rules for operation management. This mode based on static or semi-static parameter setting ignores the complex dynamic changes faced by the sewage treatment system in actual operation.

[0004] The existing technology faces many challenges in actual sewage treatment conditions, and its shortcomings are increasingly prominent. The frequent fluctuations of influent water quality and quantity, such as rainwater impact, can lead to increased water quantity and reduced pollutant concentration. At this time, the traditional cycle adjustment often involves manual intervention or rough model-based adjustment, which is difficult to achieve real-time, accurate and globally optimal self-adaptive optimization of key operation parameters such as total operation cycle, sludge reflux ratio, aeration intensity, sludge-water mixing time, etc. Such rigid control strategy is extremely easy to lead to unstable system treatment efficiency, increased carbon source consumption, excessive energy consumption or substandard effluent under complex and variable environment, thereby seriously affecting the economy, efficiency and environmental benefits of sewage treatment. Therefore, the existing technology urgently needs an intelligent control means that can actively learn environmental changes and adjust operation strategies in real time, so as to maximize the energy efficiency of the sewage treatment system and optimize the stability of the effluent. SUMMARY

[0005] The application discloses an assembled aerobic granular sludge reactor operation parameter self-adaptive method and system. The method aims to solve the problem that the existing assembled aerobic granular sludge reactor operation management generally relies on preset standard cycles, experience adjustment or simple rule control strategies. The existing mode ignores the complex dynamic changes faced by the wastewater treatment system in actual operation, resulting in unstable system treatment efficiency, increased carbon source consumption, excessive energy consumption and substandard effluent. The application realizes real-time self-adaptive adjustment of key parameters such as total operation cycle, sludge reflux ratio, aeration intensity and sludge-water mixing time, so as to maximize the energy efficiency of the wastewater treatment system and optimize the stability of the effluent, by real-time sensing of multi-dimensional operation data of the reactor, deep processing and multi-modal feature fusion, precise prediction of the future operation state of the reactor by a deep learning model, and determination of the optimal operation parameter combination by a multi-objective optimization algorithm.

[0006] According to one aspect of the application, an assembled aerobic granular sludge reactor operation parameter self-adaptive method is provided, which specifically comprises the following steps: Real-time collection of multi-dimensional operation data of the assembled aerobic granular sludge reactor, wherein the multi-dimensional operation data includes influent water quality and quantity data, reactor internal biological environment parameter data, effluent water quality data and equipment operation state data; Data preprocessing of the multi-dimensional operation data to eliminate data noise, fill in missing values and realize data standardization and alignment; Based on the preprocessed multi-dimensional operation data, a multi-modal fusion feature set is constructed to represent the comprehensive operation state and environmental dynamics of the reactor at present and in the past; The multi-modal fusion feature set is input into a pre-trained deep learning prediction model to predict the key performance indicators of the reactor within a specific time window in the future, including effluent water quality parameters and energy consumption indicators; According to the predicted key performance indicators and preset optimization objectives, a set of optimal assembled aerobic granular sludge reactor operation parameters is determined by a multi-objective optimization algorithm, including total operation cycle, sludge reflux ratio, aeration intensity and sludge-water mixing time; The optimal operation parameters are converted into specific control instructions and issued to the actuators of the assembled aerobic granular sludge reactor to realize real-time self-adaptive adjustment of the reactor operation parameters; The actual operation response data of the actuators after adjustment are continuously monitored, and the actual operation response data are fed back to the deep learning prediction model and the multi-objective optimization algorithm to iteratively update the model parameters, realize adaptive learning and performance optimization.

[0007] In accordance with another aspect of the present application, there is provided an assembled aerobic granular sludge reactor operation parameter adaptive system, which specifically comprises: a data acquisition module for acquiring multi-dimensional operation data of the assembled aerobic granular sludge reactor in real time, wherein the multi-dimensional operation data includes influent water quality and quantity data, reactor internal biological environment parameter data, effluent water quality data, and equipment operation state data; a data preprocessing module for preprocessing the multi-dimensional operation data to eliminate data noise, fill in missing values, and realize data standardization and alignment; a feature engineering module for constructing a multi-modal fusion feature set based on the preprocessed multi-dimensional operation data to represent the comprehensive operation state and environmental dynamics of the reactor at present and in the past; a deep learning prediction module for inputting the multi-modal fusion feature set into a pre-trained deep learning prediction model to predict key performance indicators of the reactor within a specific time window in the future, wherein the key performance indicators include effluent water quality parameters and energy consumption indicators; a multi-objective adaptive optimization module for determining a set of optimal assembled aerobic granular sludge reactor operation parameters, including total operation period, sludge reflux ratio, aeration intensity, and sludge-water mixing time, through a multi-objective optimization algorithm based on the predicted key performance indicators and preset optimization objectives; an operation parameter instruction generation module for converting the optimal operation parameters into specific control instructions and issuing them to the actuators of the assembled aerobic granular sludge reactor to realize real-time adaptive adjustment of the reactor operation parameters; a feedback and learning module for continuously monitoring the actual operation response data of the reactor after adjustment by the actuators and feeding the actual operation response data back to the deep learning prediction module and the multi-objective adaptive optimization module to iteratively update the model parameters, realize adaptive learning and performance optimization.

[0008] Compared with the prior art, the present application has the following advantages and positive effects: The present application solves the problem that the traditional control strategy cannot realize real-time, accurate, and globally optimal adaptive adjustment of operation parameters when the existing assembled aerobic granular sludge reactor faces influent water quality and quantity fluctuations. By constructing a multi-dimensional and multi-modal data sensing and fusion mechanism, the present application can comprehensively and accurately grasp the internal biological state, external environmental conditions, and equipment operation state of the reactor, breaking the control limitations caused by single data source or local information.

[0009] The application introduces a deep learning prediction model to realize accurate prediction of future key performance indicators of the reactor. The model can capture complex nonlinear relationships and time-dependent dependencies, enabling the control system to shift from passive response to active prediction, providing a forward-looking basis for optimization decisions and effectively avoiding water over-standard or energy waste caused by lagging response.

[0010] The application adopts a multi-objective optimization algorithm, which can comprehensively consider multiple mutually restrictive optimization objectives such as effluent water quality, energy consumption, carbon source consumption, and sludge activity and form stability, and find the best combination of operating parameters in a complex optimization space. This overcomes the drawbacks of traditional experience adjustment or single-objective optimization methods that cannot consider multiple performance indicators, achieving coordinated optimization of treatment efficiency, environmental benefits, and economic benefits.

[0011] The application realizes precise adaptive adjustment of key operating parameters such as total operating cycle, sludge reflux ratio, aeration intensity, and sludge-water mixing time by converting the optimization results into specific control instructions and linking them with the reactor actuators in real time. This fine control avoids over-aeration or insufficient aeration, reduces unnecessary carbon source addition, and thus reduces operating costs.

[0012] The application establishes a continuous feedback and learning mechanism, enabling the deep learning prediction model and multi-objective optimization algorithm to iteratively update and self-improve based on actual operating response data. This mechanism gives the system true adaptive ability, enabling it to run stably and adapt to environmental changes over a long period. As the running time increases, the optimization performance of the system gradually improves, and the treatment effect becomes more stable and reliable.

[0013] The application significantly improves the intelligence and automation level of the assembled aerobic granular sludge reactor, reduces the dependence on human experience, effectively ensures the stable compliance of effluent water quality, and significantly reduces energy consumption and carbon source consumption, providing an efficient, economical, and environmentally friendly solution for municipal and industrial wastewater treatment. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 is the overall technical scheme architecture diagram of the assembled aerobic granular sludge reactor operating parameter adaptive system proposed by the application.

[0015] Figure 2 is the core principle framework diagram of the assembled aerobic granular sludge reactor adaptive control in the application.

[0016] Figure 3 is the logic flow framework diagram of multi-dimensional operating data collection, preprocessing, and multi-modal feature construction in the application.

[0017] Figure 4is the core principle framework diagram of the deep learning prediction model and multi-objective optimization algorithm collaborative decision-making in the present application.

[0018] Figure 5 is the logic flow framework diagram of the feedback and adaptive learning mechanism in the present application. DETAILED DESCRIPTION

[0019] The embodiment provides an assembled aerobic granular sludge reactor operation parameter adaptive method and system. The core of the method is to realize real-time adaptive adjustment of key operation parameters of the reactor by comprehensive data perception, intelligent data processing and feature extraction, accurate future state prediction and multi-objective optimization decision, so that the energy efficiency is maximized and the operation cost is reduced while ensuring that the effluent meets the standard. The system serves as the physical and logical carrier supporting the method, ensuring the collaborative operation and continuous optimization of each link.

[0020] Referring to Figures 2-5 , the assembled aerobic granular sludge reactor operation parameter adaptive method specifically includes seven steps that are interrelated and cyclically iterated, so as to ensure that the system can maintain efficient and stable operation performance when facing complex and variable sewage quality and environmental conditions.

[0021] S1: Real-time collection of multi-dimensional operation data of the assembled aerobic granular sludge reactor.

[0022] This step is the basis of the entire adaptive control system, which aims to comprehensively and accurately obtain real-time information of the internal and external environment of the reactor, the inflow and outflow conditions and the equipment operation state. The range of data collection includes inflow water quality and quantity data, reactor internal biological environment parameter data, effluent quality data and equipment operation state data, forming a complete multi-dimensional operation data set.

[0023] As an embodiment of the present application, the real-time collection of multi-dimensional operation data of the assembled aerobic granular sludge reactor specifically includes the following sub-steps and specific data acquisition methods: First, influent quality and quantity monitoring: Influent flow rate, chemical oxygen demand, ammonia-nitrogen concentration, total nitrogen concentration, and total phosphorus concentration are obtained through an influent flow meter, an online chemical oxygen demand analyzer, an online ammonia-nitrogen analyzer, an online total nitrogen analyzer, and an online total phosphorus analyzer. The influent flow meter usually uses an electromagnetic flow meter, with a measurement range from 0.1 cubic meters per hour to 50 cubic meters per hour, an accuracy of 1%, and an output of 4-20 mA standard analog signals for real-time monitoring of the amount of wastewater entering the reactor. The online chemical oxygen demand analyzer uses potassium dichromate oxidation or ultraviolet absorption method, with a measurement range of 10-2,000 mg / L, an accuracy of 5%, and a monitoring period of every 15 minutes to ensure timely grasp of organic pollutant load. The online ammonia-nitrogen analyzer uses Nessler's reagent colorimetric method or ion selective electrode method, with a measurement range of 0.1-100 mg / L, an accuracy of 3%, and a monitoring period of every 30 minutes. The online total nitrogen analyzer and the online total phosphorus analyzer use ultraviolet-visible spectrophotometry, with a total nitrogen measurement range of 0.5-150 mg / L, an accuracy of 4%, a total phosphorus measurement range of 0.001-20 mg / L, an accuracy of 5%, and a monitoring period of every hour to evaluate total nutrient load.

[0024] Second, internal biological environment parameter monitoring: Dissolved oxygen concentration, oxidation-reduction potential, sludge concentration, pH value, and sludge settling ratio inside the reactor are obtained through a dissolved oxygen sensor, an oxidation-reduction potential sensor, a sludge concentration sensor, a pH sensor, and an online sludge settling ratio analyzer. The dissolved oxygen sensor uses fluorescence or electrochemical method, with a measurement range of 0-20 mg / L, an error of not more than ±0.1 mg / L, and a data update frequency of every 10 seconds for accurate control of the aeration process. The oxidation-reduction potential sensor monitors a range of -1,000 mV to +1,000 mV with an accuracy of ±5 mV, providing key information on the oxidation-reduction state inside the reactor. The sludge concentration sensor uses optical principles such as infrared scattering or transmission method, with a measurement accuracy of 2%, for real-time monitoring of mixed liquor suspended solids concentration. The pH sensor measures a range of 0-14 with an accuracy of ±0.02 pH units and a data update frequency of every minute for evaluation of biological activity and nitrification-denitrification process. The online sludge settling ratio analyzer automatically measures sludge settling performance through optical or ultrasonic methods, providing dynamic indicators of sludge granulation and settling performance.

[0025] Third, effluent quality monitoring: The chemical oxygen demand, ammonia nitrogen concentration, total nitrogen concentration, and total phosphorus concentration of the effluent are obtained through online chemical oxygen demand analyzers, online ammonia nitrogen analyzers, online total nitrogen analyzers, and online total phosphorus analyzers. The technical specifications, measurement ranges, accuracies, and monitoring periods of these sensors are consistent with the analyzers used for influent quality monitoring, ensuring accurate assessment of effluent quality to determine whether the treatment effect meets the standards.

[0026] Fourth, equipment energy consumption and operating state monitoring: The power of the air blower, the power of the water pump, the power of the agitator, the internal pressure of the reactor, and the liquid level height are obtained through electric energy meters, pressure sensors, and liquid level sensors. Electric energy meters are used to accurately measure the real-time electric power consumption of main energy-consuming equipment such as air blowers, water pumps, and agitators, with an accuracy of 0.5% and a data update frequency of once per minute. Pressure sensors are installed in the aeration system pipeline and the reactor bottom, with a measurement range of 0-50 kPa and an accuracy of 0.1%, used to monitor the operating pressure of the aeration system and the water pressure distribution inside the reactor. Liquid level sensors use ultrasonic or static pressure, with a measurement range covering the effective working water level interval of the reactor, an accuracy of millimeter level, and output signals of continuous analog or discrete digital, used to accurately control the water inlet, drainage, and decanting process of the reactor.

[0027] Fifth, environmental monitoring data collection: Environmental monitoring data are obtained through water level sensors, temperature sensors, flow rate sensors or flow meters, and environmental temperature sensors, environmental humidity sensors, atmospheric pressure sensors, and rainfall sensors. Water level sensors are used to monitor the internal liquid level of the reactor, ensuring that the water level is within the safe operating range. Temperature sensors are distributed in different areas of the reactor, including water inlet, water outlet, mixed liquid, and returned sludge, with a measurement accuracy of 0.1°C and a data sampling frequency of once per minute, used to monitor water temperature and sludge temperature. Flow rate sensors or flow meters are used to measure the water inlet, water outlet, and sludge return flow, with a measurement accuracy of 0.1%, ensuring accurate control of water balance and return ratio. Environmental temperature sensors have a measurement accuracy of 0.5°C, environmental humidity sensors have a measurement accuracy of 3% relative humidity, atmospheric pressure sensors have a measurement accuracy of 0.1 kPa, and rainfall sensors record real-time rainfall conditions with an accuracy of 0.1 mm. These environmental data are used to evaluate the impact of external environment on reactor operation, such as the impact of temperature changes on microbial activity and the impact of rainfall on influent water quality and quantity.

[0028] All the above data are synchronously collected at a configurable sampling frequency, for example, set to once every five minutes. All collected data are time-stamped, preliminarily verified and packaged by the data acquisition unit, and then transmitted to the central data processing unit through industrial Ethernet, wireless sensor network or 5G communication module. The data verification mechanism includes range check, trend anomaly detection and data integrity check. The range check is used to determine whether the data is within the preset physical or logical interval. The trend anomaly detection identifies the sharp fluctuations or unreasonable changes in the data in the short term through a sliding window or a statistical model. The data integrity check ensures that the data is not lost or damaged during transmission. For example, if a water quality parameter reading exceeds the normal operating range, the data acquisition unit will be marked and an alarm will be triggered.

[0029] S2: data preprocessing of the multi-dimensional operation data.

[0030] This step aims to transform the original, heterogeneous, possibly noisy or missing value data into a high-quality, unified format data set, providing reliable input for subsequent feature construction and model prediction. The implementation of data preprocessing significantly improves the data quality, ensuring the accuracy of subsequent analysis and the robustness of the model.

[0031] As an embodiment of the present application, the data preprocessing of the multi-dimensional operation data specifically includes the following sub-steps: First, anomaly detection and correction: anomaly detection and correction of the collected raw data. The anomaly detection uses the Isolation Forest algorithm or the Three Sigma criterion. The Isolation Forest algorithm constructs an isolation tree by randomly selecting features and splitting points, and judges the abnormality of the data points by calculating the difficulty of their isolation. For example, if a dissolved oxygen reading suddenly rises to fifty milligrams per liter, far exceeding the normal range, the Isolation Forest algorithm will identify it as an anomaly. The Three Sigma criterion assumes that the data follows a normal distribution, and identifies data points that deviate from the mean by more than three standard deviations as anomalies. The anomaly correction method includes replacing the abnormal values with the mean or median of the adjacent data, or using the estimated value based on the prediction model to replace them.

[0032] Second, missing value interpolation: interpolation of data containing missing values. Interpolation uses the Lagrange interpolation method or the cubic spline interpolation method. The Lagrange interpolation method estimates missing points by constructing a polynomial, which is suitable for scenarios where data fluctuates smoothly. The cubic spline interpolation method generates a piecewise cubic polynomial, which has better smoothness and local fitting ability, and is suitable for more complex data curves. For example, if the sludge concentration sensor fails at a certain time point, resulting in missing data, the system will estimate the missing value by interpolating the normal data points before and after it.

[0033] Third, timestamp alignment: align the timestamps of data with different sampling frequencies to form a unified time series dataset. Since different sensors may have different sampling periods, such as dissolved oxygen every ten seconds and total phosphorus every hour, data alignment is necessary. This sub-step ensures that all data points correspond accurately on the time axis by selecting a reference sampling frequency, such as every five minutes, and then downsampling (taking the average or median) for high-frequency data and upsampling (using interpolation or keeping the previous value) for low-frequency data.

[0034] Fourth, normalization: normalize the aligned data. Normalization uses the min-max normalization or Z-score standardization method to scale the data to a unified numerical range. Min-max normalization linearly scales the data to the interval of zero to one, suitable for scenarios where the feature has a clear upper and lower limit. Z-score standardization converts the data to a distribution with a mean of zero and a standard deviation of one, suitable for scenarios where the data distribution is approximately normal and the dimensionality needs to be eliminated. For example, converting the influent chemical oxygen demand data (range may be fifty to one thousand milligrams per liter) and dissolved oxygen concentration data (range is zero to twenty milligrams per liter) to similar numerical scales to prevent dimension differences from affecting model training.

[0035] Further, the data preprocessing also includes data denoising, using moving average filtering, wavelet transform denoising, or Kalman filtering to remove random noise in the data and smooth the data curve. Moving average filtering smooths noise by calculating the average value of data in a sliding window, which is simple and efficient. Wavelet transform denoising decomposes the signal into different frequency components, then thresholds or suppresses noise components, and then reconstructs the signal. Kalman filtering is an algorithm for optimal estimation of the state of a dynamic system with noise, suitable for scenarios with high real-time and accuracy requirements. For example, a dissolved oxygen sensor may have high-frequency fluctuations under stable aeration conditions, which can be effectively filtered out by Kalman filtering to obtain a more realistic dissolved oxygen trend.

[0036] S3: Based on the pre-processed multi-dimensional operation data, a multi-modal fusion feature set is constructed.

[0037] This step aims to extract deep and rich features from the pre-processed multi-dimensional time series data, and through multi-modal fusion technology, comprehensively represent the current and historical comprehensive operation state and environmental dynamics of the reactor, providing higher quality and more representative input for deep learning prediction models.

[0038] As an embodiment of the present application, the multi-modal fusion feature set based on the pre-processed multi-dimensional operation data specifically includes the following sub-steps: First, time-domain feature extraction: extract time-domain features from time series data. The time-domain features include mean, variance, standard deviation, skewness, kurtosis, moving average, and autocorrelation coefficient. Mean and standard deviation reflect the central tendency and fluctuation amplitude of data. Skewness measures the symmetry of data distribution. Kurtosis measures the steepness of data distribution. Moving average can smooth data and reveal long-term trends. Autocorrelation coefficient reflects the correlation of data at different time points, revealing the periodicity or trend of data. For example, the mean and variance of influent chemical oxygen demand in the past hour can represent the average level and fluctuation of influent organic load.

[0039] Second, frequency-domain feature extraction: extract frequency-domain features from time series data. The frequency-domain features include spectral amplitude and phase information obtained by fast Fourier transform. Fast Fourier transform converts time-domain signals to frequency-domain signals, spectral amplitude represents the intensity of different frequency components, and phase information reflects the relative time relationship of different frequency components. By analyzing frequency-domain features, we can capture periodic fluctuations or oscillation patterns in reactor operation, such as aeration period, settling period, or physiological rhythm of microbial community.

[0040] Third, multi-modal feature combination: combine the time-domain features and frequency-domain features with the original preprocessed data to form a high-dimensional multi-modal fusion feature vector. This combination ensures the comprehensiveness of the feature set, which contains both the details of the original data and the abstracted and transformed time-domain and frequency-domain information. For example, the final feature vector may include the current dissolved oxygen concentration, the mean and standard deviation of dissolved oxygen in the past five minutes, and the amplitude of specific frequencies of the dissolved oxygen signal.

[0041] Further, the multi-modal feature fusion employs a deep learning based feature fusion network. The network contains multi-head self-attention mechanism, where each attention head is responsible for capturing the relevance within or between different modalities. For the operational status feature sequence, water quality feature sequence and environmental feature sequence, they are mapped to high-dimensional space through independent embedding layers, and then fused by attention weighting. This fusion process can dynamically identify and enhance the contribution of different modalities to the prediction of reactor operating status and effluent quality. Specifically, the network first maps the multi-dimensional data from different sensors (such as water quality parameters, biological parameters, and equipment energy consumption parameters) through independent embedding layers respectively, to ensure the comparability between different modalities. These embedded feature sequences are then input into the multi-head self-attention mechanism. Each attention head calculates the query, key and value matrices in parallel, and realizes dynamic modeling of the relationship between different features and different modalities through the attention weight matrix. The size of the attention weight reflects the importance of a particular modality or feature to the current prediction task. The results of multiple attention heads are concatenated and linearly transformed to obtain a comprehensive multi-modal fusion feature vector. This mechanism allows the model to automatically learn and emphasize the feature modality that contributes most to the prediction task based on the current input data, thereby achieving more fine and accurate feature representation.

[0042] S4: input the multi-modal fusion feature set into a pre-trained deep learning prediction model to predict the key performance indicators of the reactor in a specific future time window.

[0043] This step is the key to realize the forward-looking control of the system. By utilizing the powerful pattern recognition and time series prediction ability of the deep learning model, the future operating status and performance of the reactor are estimated, providing a basis for subsequent optimization decisions.

[0044] As an embodiment of the present application, the deep learning prediction model employs a hybrid deep neural network architecture. The hybrid deep neural network architecture contains at least one convolutional neural network layer and at least one long short-term memory network layer.

[0045] The convolutional neural network layer is used to capture the local spatial correlation and patterns in the feature set. In a multi-modal feature vector, there may be local, nonlinear correlations between different types of features. For example, there may be some pattern of association between the influent chemical oxygen demand and the dissolved oxygen concentration in a short period of time. The convolutional layer can effectively extract these spatial patterns or feature combinations through sliding windows and filters.

[0046] The long short-term memory network layer is used to capture long-term dependencies and temporal dynamics in time series data. In reactor operation, the state at the current time is not only affected by the recent data, but also possibly by data from hours or even days ago. The long short-term memory network, with its gating mechanism (input gate, forget gate, output gate), can effectively solve the gradient vanishing or gradient exploding problem of traditional recurrent neural networks, thereby learning and memorizing long-term dependencies in time series and accurately capturing complex temporal dynamics.

[0047] The input of the deep learning prediction model is the multi-modal fusion feature set, and the output is the predicted chemical oxygen demand of effluent, the predicted ammonia nitrogen concentration of effluent, the predicted total nitrogen concentration of effluent, the predicted total phosphorus concentration of effluent, and the predicted system energy consumption within one to six hours in the future. The setting of the prediction time window (one to six hours) aims to reserve sufficient time for the generation and execution of control instructions.

[0048] The deep learning prediction model is trained by supervised learning using historical operation data and actual monitoring data. The mean squared error is used as the loss function in the training process, and the adaptive moment estimation optimizer is used for model parameter update. The mean squared error is a commonly used loss function in regression tasks, which measures the average of the squared differences between predicted values and true values, aiming to make the prediction results as close to the true values as possible. The adaptive moment estimation optimizer combines the advantages of momentum method and RMSProp, and can adaptively adjust the learning rate of each parameter, thereby accelerating the convergence of the model and improving the training stability.

[0049] Further, the multi-objective deep learning prediction model adopts a multi-task learning architecture to simultaneously predict multiple key future operating states and performance indicators. Multi-task learning can take advantage of the correlation between different prediction tasks by sharing part of the network layers, thereby improving the overall prediction performance. The objective function of the prediction model is a weighted sum, which is used to balance the accuracy of each prediction task. For example, when predicting the effluent water quality parameters and system energy consumption simultaneously, the effluent standard prediction can be given a higher weight according to actual needs, to ensure environmental compliance priority.

[0050] During model training, each training batch of data set {(X t ,Y t )} contains the multi-modal fusion feature set X t collected at time point t and the corresponding real key performance indicators Y t within the future time window.

[0051] The goal of the model is to learn a mapping function f, such that , represents the predicted value of the model for “real key performance indicators within the future time window”.

[0052] The loss function L can be expressed as: In this formula, , , respectively represent the predicted effluent chemical oxygen demand, effluent ammonia nitrogen concentration and system energy consumption. , , respectively represent the actual corresponding values. MSE represents the mean square error. , , and so on are the weights of each prediction task, and their values reflect the importance of each task. These weights can be preset before model training, or dynamically adjusted through meta-learning or hyperparameter optimization methods.

[0053] S5: According to the predicted key performance indicators and the preset optimization target, a set of optimal assembled aerobic granular sludge reactor operating parameters is determined through a multi-objective optimization algorithm.

[0054] This step is the core decision-making link of the adaptive method, which converts the prediction results of the deep learning model into operational operating parameters. By considering multiple conflicting optimization objectives, it ensures that the determined parameter combination can balance between processing efficiency, energy consumption and environmental benefits.

[0055] As an embodiment of the present application, the multi-objective optimization algorithm is implemented by a reinforcement learning agent. The reinforcement learning agent searches for the optimal strategy under the set optimization target through exploration and utilization mechanisms.

[0056] The observation space of the reinforcement learning agent is composed of the predicted key performance indicators, the current reactor state and the environmental parameters. The predicted key performance indicators include the predicted effluent chemical oxygen demand, the predicted effluent ammonia nitrogen concentration, the predicted effluent total nitrogen concentration, the predicted effluent total phosphorus concentration and the predicted system energy consumption. The current reactor state includes real-time dissolved oxygen concentration, sludge settling ratio, pH value and other biological environmental parameters. Environmental parameters include environmental temperature, atmospheric pressure, etc. These information constitutes the complete context for the agent to make decisions.

[0057] The action space of the reinforcement learning agent is composed of adjustable assembled aerobic granular sludge reactor operating parameters. The operating parameters include total operating period, sludge reflux ratio, aeration intensity and sludge-water mixing time.

[0058] The total operating period ranges from 180 minutes to 360 minutes, with a step size of 15 minutes, covering a reasonable time range required for microbial growth and pollutant degradation.

[0059] The sludge return ratio is in the range of 50% to 150%, with a step of 5%, to maintain a suitable sludge concentration and stability of granular sludge in the reactor.

[0060] The aeration intensity is in the range of 2 mg / L to 4 mg / L, with a step of 0.2 mg / L, or in the range of 40% to 80% of the blower power output, with a step of 5%, to ensure the oxygen demand of microorganisms and avoid energy waste caused by excessive aeration.

[0061] The sludge-water mixing time is in the range of 5 minutes to 20 minutes, with a step of 1 minute, to ensure sufficient contact between microorganisms and pollutants and promote mass transfer efficiency.

[0062] The reward function of the reinforcement learning agent considers multiple optimization objectives. The optimization objectives include minimizing effluent pollutant concentration, minimizing energy consumption per unit of treated water, minimizing carbon source consumption, and maintaining sludge activity and stability of granular sludge morphology. The reward function can be designed in the form of weighted sum or based on the Pareto optimal concept.

[0063] For example, the reward function R can be defined as: In this formula, , , , C, NH4-N, TN, and TP represent the effluent chemical oxygen demand, effluent ammonia nitrogen concentration, effluent total nitrogen concentration, and effluent total phosphorus concentration, respectively, after appropriate normalization or penalty function processing to reflect the degree of deviation from the limit value. E represents the energy consumption per unit of treated water. C represents the carbon source consumption. S represents a negative indicator of sludge activity and stability of granular sludge morphology, such as sludge disintegration rate or sludge settling performance decline. are the corresponding weight coefficients, which can be adjusted according to the priority of the operation strategy, for example, giving higher weight to water quality indicators when the effluent water quality pressure is greater. The negative sign of the reward function indicates that the optimization objective is to minimize these negative indicators.

[0064] ​​​​​​The reinforcement learning agent is trained using a deep Q-network or an asynchronous advantage actor-critic algorithm. Deep Q-network approximates the action-value function using a deep neural network, which is suitable for solving problems with discrete action spaces. The asynchronous advantage actor-critic algorithm combines policy gradient and value function learning, which is capable of handling continuous action spaces and has better convergence and stability.

[0065] Further, the multi-objective optimization problem can also be solved using evolutionary algorithms. The evolutionary algorithms include non-dominated sorting genetic algorithm or multi-objective particle swarm optimization. The non-dominated sorting genetic algorithm finds the Pareto optimal solution set by non-dominated sorting and crowding distance calculation on the population. The multi-objective particle swarm optimization searches for the optimal solution in multi-dimensional space by simulating the foraging behavior of bird flocks. The optimization objective function is a weighted sum that balances the priority of each optimization objective. The constraints include effluent quality constraints, operating parameter physical constraints, and sludge health constraints. The effluent quality constraints ensure that the effluent indicators are below the national or local emission standards. The operating parameter physical constraints ensure that the operating parameters are within the operating range allowed by the equipment. The sludge health constraints ensure that the sludge activity, particle morphology, and microbial community structure are maintained in a healthy state.

[0066] S6: Convert the optimal operating parameters into specific control instructions and issue them to the actuators of the assembled aerobic granular sludge reactor.

[0067] This step is a bridge that converts intelligent decision-making into actual action, ensuring that the optimal operating parameters generated by the optimization algorithm can be accurately executed by the automation equipment on site.

[0068] As an embodiment of the present application, the conversion of the optimal operating parameters into specific control instructions specifically includes the following sub-steps: First, periodic adjustment instruction generation: send the optimal total operating period instruction to the programmable logic controller to adjust the operating time of the periodic switching valve and pump. The programmable logic controller adjusts the start-stop sequence and duration of devices such as the influent pump, effluent pump, backflow pump, and decanter according to the received total operating period instruction, ensuring that the reactor operates periodically according to the new cycle. For example, if the total operating period is adjusted from two hundred and forty minutes to two hundred and seventy minutes, the time length of each stage such as influent, aeration, sedimentation, and decanting will be proportionally redistributed and converted into specific timer and counter parameters, which are issued to the programmable logic controller.

[0069] Second, backflow ratio control instruction generation: send the optimal sludge backflow ratio instruction to the frequency converter of the backflow sludge pump to adjust the speed of the backflow sludge pump. The backflow sludge pump frequency converter accurately adjusts the output frequency and voltage of the motor according to the received instruction, thereby changing the speed of the pump and controlling the flow of backflow sludge to achieve the preset sludge backflow ratio. For example, if the instruction requires a backflow ratio of 70%, the frequency converter will calculate the corresponding speed and drive the backflow sludge pump to run at that speed.

[0070] Third, aeration intensity control instruction generation: send the optimal aeration intensity instruction to the blower frequency converter to adjust the speed or air volume of the blower, and simultaneously link the dissolved oxygen controller to achieve precise aeration control. If the aeration intensity is given in the form of a dissolved oxygen concentration target value, the system will send this target value to the dissolved oxygen controller. The dissolved oxygen controller monitors the internal dissolved oxygen concentration of the reactor in real time and adjusts the output of the blower frequency converter through a proportional-integral-derivative control algorithm, so that the actual dissolved oxygen concentration can be stabilized around the target value. If the aeration intensity is given in the form of a blower power output percentage, the percentage instruction is directly sent to the blower frequency converter, which runs according to the specified power percentage.

[0071] Fourth, sludge-water mixing time control instruction generation: send the optimal sludge-water mixing time instruction to the mixer controller to adjust the running time or intensity of the mixer. After receiving the instruction, the mixer controller accurately controls the start-stop time or speed of the mixing motor to ensure that the sludge-water is fully mixed within the specified time, facilitating the formation of granular sludge and the progress of biological reactions.

[0072] The control instructions are communicated to the actuators on site through digital output or analog output interfaces. Digital output interfaces are usually used for start-stop control, and analog output interfaces are usually used for continuous quantity control such as adjusting the frequency converter and valve opening.

[0073] Further, the specific control instructions are packaged into standardized control command packets. The command packets follow Modbus TCP, Profibus, or industrial Ethernet protocols. These industrial communication protocols ensure the reliability, real-time performance, and compatibility of instruction transmission. The instruction delivery channel can be hard-wired or wireless communication. Hard-wired connection provides the highest stability and anti-interference capability. Wireless communication, such as through a 5G module, provides greater flexibility and deployment convenience, especially for distributed or remote control scenarios.

[0074] S7: continuously monitor the actual operation response data of the reactor after adjustment of the actuators and feed the actual operation response data back to the deep learning prediction model and multi-objective optimization algorithm to iteratively update the model parameters, achieving adaptive learning and performance optimization.

[0075] This step builds a closed-loop control system by collecting real-time adjusted operation effect, evaluating the effectiveness of the control strategy, and continuously optimizing the prediction model and decision algorithm based on this, to ensure that the system can continuously improve performance over time and environmental changes.

[0076] As an embodiment of the present application, the continuous monitoring of the actual operation response data of the reactor after adjusting the actuator, and feeding the actual operation response data back to the deep learning prediction model and multi-objective optimization algorithm to iteratively update the model parameters specifically includes the following sub-steps: First, actual response data collection: collect the actual effluent water quality parameters, energy consumption data, sludge characteristic parameters and other operation state data of the reactor after the execution of the control instruction. These data are similar to the data collected in S1 step, but the collection time point is immediately after the execution of the control instruction, which is used to evaluate the adjustment effect. For example, the system will collect the actual dissolved oxygen concentration of the reactor after adjusting the aeration intensity, as well as the chemical oxygen demand and energy consumption data that follow.

[0077] Second, model correction and update: compare the actual operation response data with the prediction results of the deep learning prediction model, calculate the prediction error, and use the error for incremental learning or periodic retraining of the model. Incremental learning allows the model to update its parameters gradually when receiving new data without starting from scratch, which is suitable for scenarios with high real-time requirements. Periodic retraining uses accumulated new and old data to retrain the model at certain time intervals to ensure that the model can adapt to long-term trend changes. For example, if the prediction model underestimates the chemical oxygen demand of the effluent at a certain time, the deviation between the actual data and the prediction result will be used to adjust the weights and biases of the model to reduce future prediction errors.

[0078] Third, strategy optimization and adjustment: use the actual operation response data as new experience samples of the reinforcement learning agent to update the policy network parameters of the reinforcement learning agent in the multi-objective adaptive optimization module. The reinforcement learning agent compares the actual operation effect with the expected reward, if the actual effect is better than expected, the strategy that leads to this behavior is enhanced; if the effect is not good, the strategy is weakened, so that its policy network parameters converge to the optimal direction. This continuous trial and error and learning mechanism enables the optimization algorithm to adapt to long-term changes in reactor operating conditions and improve the accuracy and robustness of decision-making.

[0079] Further, the performance evaluation mechanism also includes calculating the effluent compliance rate, energy efficiency, and sludge activity indicators. The effluent compliance rate reflects the degree to which the system meets environmental discharge standards. The energy efficiency is measured by the energy consumption per unit of treated water. The sludge activity indicators, such as respiratory rate or dehydrogenase activity, assess the physiological state of microorganisms. The model retraining and parameter updating also include adding new historical operation data to the training data set for updating the model's weights and biases. The self-adaptation of the strategy optimization algorithm also includes adjusting the weight parameters or search strategy of the multi-objective optimization algorithm when it is found that the actual effect of the adaptive adjustment strategy fails to achieve the optimization goal. For example, if it is found that the effluent total nitrogen concentration is consistently high during long-term operation, the system can automatically adjust the weight of total nitrogen in the multi-objective optimization algorithm to give it higher priority in the optimization process, or adjust the balance strategy of exploration and utilization of reinforcement learning to more actively explore new parameter combinations.

[0080] Referring to Figure 1 The present application also provides an assembled aerobic granular sludge reactor operation parameter adaptive system. The system is a physical and logical carrier for implementing the above method, composed of a series of interconnected and cooperative modules, which collectively support the entire process of multi-dimensional data collection, intelligent analysis, predictive decision-making, parameter regulation, and feedback learning.

[0081] The specific composition of the assembled aerobic granular sludge reactor operation parameter adaptive system includes the following core modules: One, data collection module: used for real-time collection of multi-dimensional operation data of the assembled aerobic granular sludge reactor. The multi-dimensional operation data includes influent water quality and quantity data, reactor internal biological environment parameter data, effluent water quality data, and equipment operation state data.

[0082] As an embodiment of the present application, the data collection module specifically includes: Influent water quality and quantity monitoring subunit, including influent flow meter, online chemical oxygen demand analyzer, online ammonia nitrogen analyzer, online total nitrogen analyzer, and online total phosphorus analyzer, for obtaining influent flow, influent chemical oxygen demand, influent ammonia nitrogen concentration, influent total nitrogen concentration, and influent total phosphorus concentration. These sensors and analyzers are connected to the central processor of the data collection module through standard interfaces for data reading and preliminary verification.

[0083] Reactor internal state monitoring subunit, including dissolved oxygen sensor, oxidation-reduction potential sensor, sludge concentration sensor, pH sensor, and sludge settling ratio online analyzer, for obtaining the dissolved oxygen concentration, oxidation-reduction potential, sludge concentration, pH value, and sludge settling ratio inside the reactor. These sensors usually use analog or digital signal output and are connected through the input port of the data collection module.

[0084] Effluent water quality monitoring subunit, which contains online chemical oxygen demand analyzer, online ammonia nitrogen analyzer, online total nitrogen analyzer and online total phosphorus analyzer, is used to obtain the effluent chemical oxygen demand, effluent ammonia nitrogen concentration, effluent total nitrogen concentration and effluent total phosphorus concentration. The layout and function of these analyzers are similar to the influent water quality monitoring subunit, but are installed in the reactor effluent pipeline.

[0085] Device energy consumption and operating state monitoring subunit, which contains electric energy meter, pressure sensor and liquid level sensor, is used to obtain the power of the air blower, the power of the water pump, the power of the agitator, the internal pressure of the reactor and the liquid level height. The electric energy meter monitors the power through the current transformer and the voltage transformer, and the pressure sensor and the liquid level sensor provide analog or digital signals.

[0086] All monitoring subunits are connected with the data processing unit through industrial Ethernet or RS485 bus, and the data is synchronously collected at a uniform sampling frequency. Industrial Ethernet provides high-speed and reliable data transmission, and RS485 bus is suitable for long-distance and multi-point communication.

[0087] Further, the data acquisition module also includes water level sensor, temperature sensor, flow rate sensor or flow meter, and environmental temperature sensor, environmental humidity sensor, atmospheric pressure sensor and rainfall sensor, which are used to obtain the liquid level, water temperature, sludge temperature, influent and effluent water and sludge return flow, and environmental temperature, humidity, atmospheric pressure and rainfall data. These auxiliary sensors further enrich the data dimension and improve the system's perception ability to environmental changes.

[0088] Second, the data preprocessing module: for data preprocessing of the multi-dimensional operating data, to eliminate data noise, fill in missing values and realize data standardization and alignment.

[0089] As an embodiment of the present application, the data preprocessing module specifically includes: The outlier processing unit is used to detect and correct outliers in the original data by the Isolation Forest algorithm or the three-sigma criterion. This unit receives the original data stream from the data acquisition module and executes the outlier detection algorithm in real time.

[0090] The missing value interpolation unit is used to supplement the data containing missing values by the Lagrange interpolation method or the cubic spline interpolation method. This unit interpolates the missing or corrected outliers after the outlier processing.

[0091] Data alignment unit for timestamp alignment of data with different sampling frequencies to form a unified time series dataset. This unit is responsible for unifying data streams from different sensors and analyzers to a common time reference.

[0092] Data normalization unit for scaling the aligned data to a unified numerical range through min-max normalization or Z-score standardization methods. This unit performs the final numerical scale adjustment on the aligned data.

[0093] Further, the data preprocessing module also includes a data denoising unit for removing random noise in the data through methods such as moving average filtering, wavelet transform denoising, or Kalman filtering. This unit is usually executed before or after standardization to provide smoother and more stable data.

[0094] Three, feature engineering module: for constructing a multi-modal fusion feature set based on pre-processed multi-dimensional operation data to represent the current and historical comprehensive operation state and environmental dynamics of the reactor.

[0095] As an embodiment of the present application, the feature engineering module specifically includes: Time domain feature extraction unit for extracting mean, variance, standard deviation, skewness, kurtosis, moving average, and autocorrelation coefficient from time series data. This unit applies statistical analysis methods to pre-processed data.

[0096] Frequency domain feature extraction unit for obtaining frequency spectrum amplitude and phase information of time series data through fast Fourier transform. This unit uses signal processing techniques to mine periodic patterns from data.

[0097] Multi-modal feature fusion unit for combining the time domain features, frequency domain features, and original pre-processed data to form a high-dimensional multi-modal fusion feature vector. This unit integrates different types of features into a unified input vector.

[0098] Further, the multi-modal feature fusion unit adopts a deep learning-based feature fusion network containing a multi-head self-attention mechanism. This deep learning network can automatically learn and optimize the combination of features, improving the expression ability of features.

[0099] Four, deep learning prediction module: for inputting the multi-modal fusion feature set into a pre-trained deep learning prediction model to predict key performance indicators of the reactor in a specific future time window.

[0100] As an embodiment of the present application, the deep learning prediction module specifically includes: a hybrid deep neural network model comprising at least one convolutional neural network layer and at least one long short-term memory network layer. The convolutional neural network layer is used to capture local spatial correlation, and the long short-term memory network layer is used to capture time series dependency.

[0101] a model training unit configured to perform supervised learning training on the hybrid deep neural network model by using historical operation data and actual monitoring data. The training process uses mean square error as a loss function and an adaptive moment estimation optimizer.

[0102] The input interface of the hybrid deep neural network model receives the multi-modal fusion feature set, and the output interface outputs predicted chemical oxygen demand of effluent, predicted ammonia nitrogen concentration of effluent, predicted total nitrogen concentration of effluent, predicted total phosphorus concentration of effluent, and predicted system energy consumption within one to six hours in the future.

[0103] Five, a multi-objective adaptive optimization module: configured to determine a set of optimal assembly type aerobic granular sludge reactor operating parameters according to the predicted key performance indicators and preset optimization objectives by using a multi-objective optimization algorithm.

[0104] As an embodiment of the present application, the multi-objective adaptive optimization module specifically comprises: a reinforcement learning agent configured to search for an optimal strategy under the set optimization objectives by using an exploration and exploitation mechanism.

[0105] a state observation unit configured to receive the predicted key performance indicators output by the deep learning prediction module, the current reactor state, and the environmental parameters, and construct an observation space of the reinforcement learning agent.

[0106] an action decision unit configured to determine optimal operating parameters in an action space composed of total operating period, sludge reflux ratio, aeration intensity, and sludge-water mixing time according to the current strategy of the reinforcement learning agent.

[0107] a reward calculation unit configured to calculate a reward value of the reinforcement learning agent according to actual operation feedback data and preset optimization objectives. The reward value comprehensively considers effluent pollutant concentration, energy consumption per unit of treated water, carbon source consumption, and sludge activity and granular sludge form stability.

[0108] a strategy updating unit configured to iteratively update strategy network parameters of the reinforcement learning agent according to the reward value by using a deep Q network or an asynchronous advantage actor critic algorithm.

[0109] Further, the multi-objective adaptive optimization module further comprises an evolutionary optimization unit configured to solve a multi-objective optimization problem by using a non-dominated sorting genetic algorithm or a multi-objective particle swarm optimization algorithm.

[0110] Six, operation parameter instruction generation module: for converting the optimal operation parameters into specific control instructions and issuing them to the actuators of the assembled aerobic granular sludge reactor.

[0111] As an embodiment of the present application, the operation parameter instruction generation module specifically comprises: Period adjustment instruction generation unit, for converting the optimal total operation period into an adjustment instruction recognizable by the programmable logic controller.

[0112] Backflow proportion control instruction generation unit, for converting the optimal sludge backflow proportion into a speed instruction of the backflow sludge pump frequency converter.

[0113] Aeration intensity control instruction generation unit, for converting the optimal aeration intensity into a speed or air volume instruction of the blower frequency converter, and linking the dissolved oxygen controller.

[0114] Sludge-water mixing time control instruction generation unit, for converting the optimal sludge-water mixing time into an operation time or intensity instruction of the agitator controller.

[0115] The operation parameter instruction generation module communicates with the programmable logic controller or distributed control system on site of the assembled aerobic granular sludge reactor through a digital output or analog output interface.

[0116] Further, the instruction generation unit further comprises an instruction packaging unit for packaging the specific control instructions into standard control command packets complying with Modbus TCP, Profibus or industrial Ethernet protocols.

[0117] Seven, feedback and learning module: for continuously monitoring the actual operation response data of the reactor after adjustment of the actuators, and feeding the actual operation response data back to the deep learning prediction module and the multi-objective adaptive optimization module to iteratively update the model parameters, realize adaptive learning and performance optimization.

[0118] As an embodiment of the present application, the feedback and learning module specifically comprises: Actual response data acquisition unit, for collecting the effluent quality parameters, energy consumption data, sludge characteristic parameters and other operation state data actually achieved by the reactor after execution of the control instructions.

[0119] Model correction unit, for comparing the actual operation response data with the prediction results of the deep learning prediction module, calculating the prediction error, and using the error to perform incremental learning or periodic retraining on the deep learning prediction model.

[0120] A policy optimization unit is configured to use the actual operation response data as new experience samples of the reinforcement learning agent to update the policy network parameters of the reinforcement learning agent in the multi-objective adaptive optimization module.

[0121] Further, the feedback and learning module further comprises a performance evaluation unit configured to calculate the water standard compliance rate, energy efficiency, and sludge activity indicators.

[0122] Further, the system further comprises a user interaction and monitoring module. The module is a bridge connecting the operators and the intelligent control system, providing intuitive visual interfaces and necessary intervention capabilities.

[0123] The user interaction and monitoring module is configured to provide a graphical user interface for the operators to monitor the reactor operation status, water quality data, energy consumption data, environmental data, and current operation parameters in real time. The graphical user interface usually uses a high-resolution display to display real-time curves, tables, and status indicator lights of the data. The operators can view the real-time data of the influent and effluent water quality through the interface, including the concentration trend of chemical oxygen demand, ammonia nitrogen, total nitrogen, and total phosphorus. The energy consumption data shows the real-time power consumption and cumulative energy consumption of devices such as blowers, water pumps, and mixers. The environmental data shows the temperature, humidity, atmospheric pressure, and rainfall around the reactor. The current operation parameters show the total running period, sludge reflux ratio, aeration intensity, and setting values of the sludge-water mixing time.

[0124] The user interaction and monitoring module allows the operators to view historical data trends, adjust the weights of optimization objectives, and receive abnormal alarm information. The historical data trend analysis function allows the operators to view data from hours, days, or even months ago to analyze the operation mode and fault diagnosis. The adjustment function of the optimization objective weight allows flexible adjustment of the priority of each objective in the multi-objective optimization algorithm according to actual management needs, such as focusing more on the effluent water quality standard in the summer peak period and focusing more on energy minimization in the off-peak period. Abnormal alarm information is timely notified to the operators in the form of sound and light alarms, short messages, or emails, such as when a water quality parameter exceeds the upper limit or a device fails.

[0125] The user interaction and monitoring module also provides a manual intervention interface to allow the operators to make emergency parameter adjustments or system start-stop control when necessary. The manual intervention interface allows the operators to override the automatic decisions of the intelligent system and directly input new operation parameters, such as manual adjustment under extreme influent shock load or stopping part of the device operation during maintenance.

[0126] The user interaction and monitoring module exchanges data and transmits instructions with each of the above core modules through an internal data bus or network interface. The data bus ensures high-speed and reliable communication between modules, enabling the monitoring interface to refresh data in real time and transmit the operator's instructions to the corresponding control module in a timely manner. This close integration ensures that the entire adaptive system can operate autonomously while also having the necessary human supervision and intervention capabilities.

[0127] The above merely describes specific embodiments of the present application, but the technical features of the present application are not limited thereto. Any simple change, equivalent replacement or modification made on the basis of the present application to solve the basically same technical problem and achieve the basically same technical effect shall be covered within the protection scope of the present application.

Claims

1. An adaptive method for operating parameters of a prefabricated aerobic granular sludge reactor, characterized in that, include: S1. Real-time acquisition of multi-dimensional operation data of the assembled aerobic granular sludge reactor, including influent water quality and quantity data, reactor internal biological environment parameter data, effluent water quality data, and equipment operation status data; S2. Perform data preprocessing on the multi-dimensional operational data to eliminate data noise, fill in missing values, and achieve data standardization and alignment; S3. Based on the preprocessed multi-dimensional operational data, construct a multi-modal fusion feature set to characterize the current and historical comprehensive operational status and environmental dynamics of the reactor; S4. Input the multimodal fusion feature set into a pre-trained deep learning prediction model to predict the key performance indicators of the reactor within a specific future time window. The key performance indicators include effluent water quality parameters and energy consumption indicators. S5. Based on the predicted key performance indicators and the preset optimization objectives, a set of optimal operating parameters for the prefabricated aerobic granular sludge reactor are determined through a multi-objective optimization algorithm. The operating parameters include the total operating cycle, sludge return ratio, aeration intensity, and sludge-water mixing time. S6. The optimal operating parameters are converted into specific control commands and sent to the actuator of the assembled aerobic granular sludge reactor to achieve real-time adaptive adjustment of the reactor operating parameters. S7. Continuously monitor the actual operating response data of the reactor after the actuator is adjusted, and feed the actual operating response data back to the deep learning prediction model and the multi-objective optimization algorithm to iteratively update the model parameters and achieve adaptive learning and performance optimization.

2. The method according to claim 1, characterized in that, The real-time acquisition of multi-dimensional operational data from the prefabricated aerobic granular sludge reactor specifically includes: The influent flow rate, influent chemical oxygen demand, influent ammonia nitrogen concentration, influent total nitrogen concentration, and influent total phosphorus concentration are obtained through an influent flow meter, an online chemical oxygen demand analyzer, an online ammonia nitrogen analyzer, an online total nitrogen analyzer, and an online total phosphorus analyzer. Dissolved oxygen concentration, oxidation-reduction potential, sludge concentration, pH value, and sludge settling ratio inside the reactor are obtained using dissolved oxygen sensors, oxidation-reduction potential sensors, sludge concentration sensors, pH sensors, and an online sludge settling ratio analyzer. The chemical oxygen demand, ammonia nitrogen, total nitrogen, and total phosphorus concentrations of the effluent were obtained using an online chemical oxygen demand analyzer, an online ammonia nitrogen analyzer, an online total nitrogen analyzer, and an online total phosphorus analyzer. The power of the blower, the power of the water pump, the power of the agitator, the internal pressure of the reactor, and the liquid level are obtained through an electricity meter, a pressure sensor, and a liquid level sensor.

3. The method according to claim 2, characterized in that, The data acquisition also specifically includes monitoring the liquid level inside the reactor through a water level sensor, monitoring the water temperature and sludge temperature through a temperature sensor, measuring the influent, effluent and sludge return flow through a flow velocity sensor or flow meter, and acquiring environmental monitoring data through an ambient temperature sensor, an ambient humidity sensor, an atmospheric pressure sensor and a rainfall sensor.

4. The method according to claim 1, characterized in that, The data preprocessing of multi-dimensional operational data specifically includes: Outlier detection and correction are performed on the collected raw data. The outlier detection uses either the isolated forest algorithm or the three sigma criterion. Imputation is performed on data containing missing values, using Lagrange interpolation or cubic spline interpolation. Timestamp alignment is performed on data from different sampling frequencies to form a unified time series dataset; The aligned data is then standardized using either min-max normalization or Z-score normalization to scale the data to a uniform numerical range.

5. The method according to claim 4, characterized in that, The data preprocessing also includes data denoising, which uses moving average filtering, wavelet transform denoising, or Kalman filtering to remove random noise from the data.

6. The method according to claim 1, characterized in that, The construction of the multimodal fusion feature set based on the preprocessed multidimensional operational data specifically includes: Extracting time-domain features from time series data, the time-domain features include mean, variance, standard deviation, skewness, kurtosis, moving average, and autocorrelation coefficient; Frequency domain features are extracted from time series data, and the frequency domain features include spectral amplitude and phase information obtained by fast Fourier transform; The time-domain features, frequency-domain features, and original preprocessed data are combined to form a high-dimensional multimodal fusion feature vector.

7. The method according to claim 6, characterized in that, The multimodal feature fusion employs a deep learning-based feature fusion network, which includes a multi-head self-attention mechanism, where each attention head is responsible for capturing the correlations within or between different modalities.

8. The method according to claim 1, characterized in that, The deep learning prediction model adopts a hybrid deep neural network architecture, which includes at least one convolutional neural network layer and at least one long short-term memory network layer. The convolutional neural network layer is used to capture local spatial correlations and patterns in the feature set, and the long short-term memory network layer is used to capture long-term dependencies and temporal dynamics in time series data.

9. The method according to claim 1, characterized in that, The multi-objective optimization algorithm is implemented using a reinforcement learning agent, which searches for the optimal strategy under the set optimization objective through exploration and utilization mechanisms. The observation space of the reinforcement learning agent consists of the predicted key performance indicators, the current reactor state, and environmental parameters. The action space of the reinforcement learning agent consists of adjustable operating parameters of the prefabricated aerobic granular sludge reactor. These operating parameters include the total operating cycle, sludge return ratio, aeration intensity, and sludge-water mixing time. The total operating cycle ranges from 180 to 360 minutes, the sludge return ratio ranges from 50% to 150%, the aeration intensity range is achieved by setting a target dissolved oxygen concentration of 2 mg / L to 4 mg / L, or by achieving 40% to 80% of the blower power output, and the sludge-water mixing time ranges from 5 to 20 minutes.

10. An adaptive operating parameter system for a prefabricated aerobic granular sludge reactor, characterized in that, include: The data acquisition module is used to collect multi-dimensional operating data of the assembled aerobic granular sludge reactor in real time. The multi-dimensional operating data includes influent water quality and quantity data, reactor internal biological environment parameter data, effluent water quality data, and equipment operating status data. The data preprocessing module is used to preprocess the multi-dimensional operational data to eliminate data noise, fill in missing values, and achieve data standardization and alignment. The feature engineering module is used to construct a multimodal fusion feature set based on preprocessed multidimensional operational data to characterize the current and historical comprehensive operational status and environmental dynamics of the reactor. The deep learning prediction module is used to input the multimodal fusion feature set into a pre-trained deep learning prediction model to predict the key performance indicators of the reactor within a specific time window in the future. The key performance indicators include effluent water quality parameters and energy consumption indicators. The multi-objective adaptive optimization module is used to determine a set of optimal operating parameters for the prefabricated aerobic granular sludge reactor based on the predicted key performance indicators and preset optimization objectives through a multi-objective optimization algorithm. The operating parameters include the total operating cycle, sludge return ratio, aeration intensity, and sludge-water mixing time. The operating parameter instruction generation module is used to convert the optimal operating parameters into specific control instructions and send them to the actuator of the assembled aerobic granular sludge reactor to realize real-time adaptive adjustment of the reactor operating parameters. The feedback and learning module is used to continuously monitor the actual operating response data of the reactor after the actuator is adjusted, and feed the actual operating response data back to the deep learning prediction module and the multi-objective adaptive optimization module to iteratively update the model parameters and realize adaptive learning and performance optimization.

Images