Data Enhancement Method For Smart Operation Of Wastewater Treatment Process

Currently,digital automatic operation is the mainstream of the wastewater treatment process management,and data-driven smart operation supported by a massive amount of high-quality data is its future development trend.However,the wastewater treatment process data collected from the field often has quality problems,primarily in the following two aspects: one is the abnormal missing of data caused by sensor failures,network interruptions,etc.,and the other is the common use of conservative control strategies enable the data collected with high energy consumption bias.These biased lowvalue data can provide limited information to the smart operation,which may interfere the normal data feature expression,bring errors to the modeling and analysis,even introduce the high energy consumption state into the subsequent operation.Based on the demand for green and smart operation of wastewater treatment enterprises,this thesis constructs data enhancement methods for data anomaly repair and data distribution bias correction respectively without changing the actual wastewater treatment operating conditions,and the main work was summarized in the following three aspects:(1)The mathematical scenario was established by abstracting the actual wastewater treatment process,a hybrid neural network based on deep learning was proposed to predict the treated water quality and it was used as a benchmark model for subsequent data enhancement tests.The window setting was introduced in the model to determine the time span of treatment states that would interact with others during the process,and the chemical oxygen demand concentration was predicted accordingly.Experiments on real dataset were conducted to determine the appropriate window lengths and training set proportions,and the results were discussed along with the treatment states represented by the data,eliciting the problem of anomalous missing and distribution bias in the wastewater treatment data.(2)For the non-regularized missing situations in wastewater treatment process data,an adaptive filling framework with embedding model was designed.The framework first performs a row-by-row search for the missing in the input dataset,marks the specific number and location of the missing after the identification.Then,it adaptively adjusts the structure of the embedded time-series model according to the specific missing cases and optimizes the model parameters by historical data,whereby the values of the missing locations are estimated and recovered.The accuracy of the proposed method was verified on the experimental dataset in combination with the comparison models.The method was applied to the entire dataset and tested using the benchmark model,and it showed that the results after anomaly recovery enhancement were significantly better than the original.This study demonstrates the destructiveness of data anomalies to what is expressed in the real and the necessity of repairing and enhancing them,and provides a universal and efficient method for missing data restoration.(3)To address the problem of high energy consumption and low-value density of data in the wastewater treatment,this thesis firstly introduced the generative adversarial network(GAN)into the wastewater treatment field and established a high-value data generation model based on GAN.After the abnormal data restoration,the real data distribution was obtained by water quality parameter analysis,and the mapping from the known data distribution to target distribution was completed by the adversarial training.Next,the data was sampled from the known distribution and input to the trained model for data generation.Some low-value data in the original dataset was replaced by the generated high-value data to improve the overall value and achieve data re-enhancement.Two variables in the model were analyzed by three sets of experiments,which determined that the quantity of generated data has a greater impact on the quality of the generated data.The results showed that the proposed model was more effective in generating data under the same experimental conditions by comparing with the variable auto-encoder.Experimental results on benchmark model before and after data enhancement were compared to confirm that the proposed method can correct the data distribution bias and improve the data-driven model fitness.This study investigates the feasibility and effectiveness for introducing GAN into wastewater treatment for low-energy data generation,and provides an optimization idea from the perspective of data for smart operation of wastewater treatment process.