Artificial Neural Network Model Of Sludge Anaerobic Digestion

In recent years, sludge production increased increasingly in China. However,owing to the poor technology in treating sludge, a large part of this sludge could notbe disposed safely, and it has become an urgent environmental issue to address.Anaerobic digestion can not only realize sludge innocuity, stabilization and reduction,but also produce biogas, a kind of resource. For this reason, sludge anaerobicdigestion has been widely employed in many countries, and the most commonly usedmethod was mesophilic digestion.Sludge anaerobic digestion systems are usually designed by stable condition, namely, assuming that the loads are constant. In the actual project, compositions andconcentrations of organics in sludge are changed over time, however. Digestionprocess is affected by many factors, and there is no certain model between thesefactors and digestion output. Most of the digestion systems are designed and operatedbased on experience. So, without proper control, worse, or even corruptedperformance of these digestion systems would be gotten. Ensuring appropriateoperation conditions using traditional methods needs a large number of experiments,which will consume huge amounts of labor power, material and financial resources. Ifa model that can accurately predict the performance according to substrate andoperation conditions of digestion system is able to be built, measures could bebeforehand taken to create optimal reaction conditions to make sure the system runssteadily and efficiently. It is very expensive to establish mechanism model, andbuilding artificial neural network model is comparatively simple.In this study, based on experiment of sludge mesophilic anaerobic digestion, adynamic relationship between volatile suspended solid (VSS) of inflow sludge, pHand alkalinity of sludge and daily yield of biogas was built. To be specific, a BPneural network，the most widely used kind of neural network currently, was built byutilizing its abilities of self-learning and nonlinear mapping. In the meantime, aimingat the shortcoming of neural network with standard BP algorithm, BP neural networksusing two kinds of representative improved algorithm, momentum factor-adaptedlearning rate algorithm and Levenberg-Marquart algorithm, were put forward. Then,these three kinds of networks were trained and used to simulate, and the results ofprediction for daily yield of biogas were compared. Results of training indicated that both the two kinds of improved algorithmnetworks were possessed of great learning ability. Correlation coefficients betweentheir trained network outputs and actual values were0.979and0.980respectively,which showed their learning abilities were both stronger than standard BP algorithmneural network (correlation coefficient was0.883). Besides, results of simulationindicated that both the two kinds of improved algorithm networks were stronger thanstandard BP algorithm network in generalization ability. The network withLevenberg-Marquart algorithm gained the best predictive effect. Among the datagotten from Levenberg-Marquart algorithm network prediction, most were very closeto practical measured data, and only a very few ones were great different frommeasured ones, showing good generalization and recognition ability of this neuralnetwork. In the end, by utilizing the prediction ability of BP neural network withLevenberg-Marquart algorithm, adjusting pH and alkalinity to appropriate valuesunder different VSS quantity, the daily yields of biogas obtained significantincreasement, which further demonstrated that the network possessed preferableprediction ability and practicability.