Combustion System Modeling And Optimization Based On Deep Learning

Coal is an important part of our country's energy structure,electricity is the major driving force to promote the development of national economic,and thermal power accounts for the main component of power production.But from the view of energy,coal is the valuable non-renewable resources,and is dwindling;On the side of environment,environment pollution is worsening,attracting public's attention,the government begin to advocate energy conservation and emissions reduction,and strengthen supervision of thermal power plant;In terms of market,the power system policy force all the power plant to face with the fierce competition of increasing economic efficiency.In the premise of guarantee the safe and stable operation of coal-fired boiler,how to reduce emissions,reduce coal consumption,improve the efficiency of the boiler become a common concerned problem.Combustion system model established with traditional identification method,generally rely on hot spot test data,and need to classify different working condition to build different physi-cal model.such models are not conducive to state switching,and can't fit the slowly physical properties change with time.Currently,there are many data-driven machine learning methods applied in establishing combustion system models,but mostly have the following defects affecting the practicality:firstly,the experiment data sets is too small,algorithm will encounter server overfitting;sec-ondly,single time step input corresponding to single time step output,model can't learn from history;and finally,the selected input/output variables are too few to learn the real operation characteristics,and can't handle the multi-variable coupling problem.In this paper,depending on the the needs of combustion optimization and deficiency of existing modeling methods,using the real power plant historical DCS data recorded every 15 minutes within a month,combustion system model is builded based on time series deep learning theory,and combustion system is optimized using Adaptive Dynamic Programming.1.Proposed a type of deep LSTM model with 21 control variables and 6 state variables as combustion system predict model.Based on in-depth analysis of plant combustion system,variables characteristics and dependencies between them,determined the input and output of system model.Applied the LSTM model which has achieved outstanding performance in time series field in coal-fired boiler modeling,to take the advantage of its strong information memory ability.2.Built two predict models for comparison:multi-layer perceptron with nonlinear autore-gression using extension input.and discussed the control variables and state variables time series are coupled or not.Then simulated the three models and analysed the simu-lation result.Experimental results show that the deep LSTM model converges faster and the average prediction error decreased by 2 to 4 times than the other two models.3.Built a LSTM-HDP optimization model using HDP and deep LSTM predict model,de-signed corresponding action network and critic network.this model achieved the purpose of optimizing the five state variables,through adjusting the 19 out of 21 control variables in constraint ranges,and ensure the main steam pressure within safe limits.4.Built two optimization models for comparison:using LSTM predict model and Gradient descent optimization algorithm,named as LSTM-GD;and using decoupling NARX pre-dict model and HDP optimization algorithms,named as NARX-HDP.Experiments show that,LSTM-HDP model ensures the main steam pressure within safe limits and boiler ef-ficiency only drops by 0.2 percentage point,can stable the main steam temperature at 567?,and optimizing heat rate fell 3%,unit coal consumption fell 2.93%,flue gas emis-sions fell 6.73%.And compared with the the other two model LSTM-HDP's convergence process is the most stable and fast.