Uranium Concentration Prediction for In-situ Leaching Based on Convolution and Long Short-term Memory Networks
This paper presents a new method for predicting uranium concentration in leachate from in-situ leaching units by integrating empirical mode decomposition(EMD),convolutional neural network(CNN),long short-term memory(LSTM)and Fourier transform.The method decomposes the time series of leachate uranium concentration monitoring values into trend,periodic and random terms by EMD.By constructing CNN+LSTM neural network and combining Fourier transform and polynomial fitting to predict the trend terms,periodic terms and random terms of uranium concentration,the sum of the three predictions act as the result of uranium concentration prediction.The empirical results show that:1)EMD can effectively decompose the uranium concentration time series,and the model fit is over 50％better than the model without EMD decomposition;2)The integrated method based on EMD,CNN+LSTM,and Fourier transform has good prediction accuracy,with an average absolute percentage error(MAPE)of 0.348,which is the highest improvement of over 80％compared to models such as LSTM,back propagation(BP)and Gate Recurrent Unit(GRU).The integrated method proposed in this paper can accurately predict the uranium concentration variation in the leaching unit,solving the problem that the original method and model cannot accurately predict the nonlinear,non-stationary uranium concentration series,thus providing technical support for production planning of ground leaching mines and it helps to enhance the digitization and informatization level of China's uranium industry.
国家科技期刊平台
NSTL
万方数据
