Extreme Learning Machine Parallel Algorithm And Its Application In Prediction Of NO_x Emission

Extreme Learning Machine(ELM) is widely used in data mining for its high learning speed, less training parameters, good generalization performance, and so on. Similar to the traditional learning algorithm, ELM needs a long training period and even may cannot run for the limitation of computer memory when dealing with the mass data. To solve above problems, this paper comes up with the idea that is to couple ELM with current popular big data technique and use the MapRuduce parallel framework to realize the parallelism of extreme learning machine and finally achieve the purpose of high-activity processing of mass data.Energy consumption increases year by year with the rapid development of economy. Coal constructs the main energy structure in our country, and NO_x produced by the coal-fired power plants has become increasingly serious. As a result, how to control the NO_x emission effectively has always been the focus of attention. With the bringing forward of industry 4.0 and the improvement of storage technology, industry has entered into a data-outbreak era. New data is generated all the time and the database storage also increases day by day. In order to cope with the mass date emerging in the informatization construction of our country’s thermal power industry, looking for the effective mass data processing technology has been a hot research.The main points of this paper are as follows:1. Set up of parallelism integrated method of extreme learning machine under MapReduce parallel framework to cope with the problem of traditional extreme learning machine which cannot deal with mass date effectively. Through the analysis of the integrated method of extreme learning machine, it can be found that the training of each extreme learning machine could run dependently and in parallel and to prove that this process can be decomposed. This algorithm is realized by a MapReduce Job and the Map stage trains each integrated ELM network in parallel. Experiments show that this algorithm owns almost the same accuracy with the single ELM integration method, but it can deal with mass data more quickly. In a word, this algorithm can be more efficient in processing massive amounts of data with good accuracy and speed.2. Put forward parallel particle swarm ELM method based on MapReduce parallel framework to cope with the problem of non-optimal input weights and hidden layer bias issue. Researches reveal that the calculation of particles fitness value depends on the size of the sample during operation process of particle swarm ELM, for example, if the sample is too large, the training time will be rather long. Fitness function uses an error of ELM network. When ELM matrix solver dealing with the largest part of the calculation, it is more effective because it can be broken down and parallelized. As the experimental results show, compared with the ELM, the algorithm has higher prediction accuracy.3. Apply parallel ELM algorithm to the prediction of NO_x emission. First of all, obtain the input and output parameters of ELM by the analysis of the boiler operating parameters and the parameters which impact the emission characteristics of NO_x. Secondly, establish NO_x emission prediction model. Finally, the validity of this method is verified by experiments.