Time series forecasting of precipitation patterns over Lucknow region using LSTM

Rainfall forecasting has assumed an important role in recent times due to uncertainities emanating from climate change as a result of environmental phenomenons like El Nino, La Nina, global warming, etc. Agriculture in India is still pretty much dependent upon rains, more so in a state like Uttar Pradesh. So it is imperative that forecasting systems are developed that can analyse the previous trends of rainfall and predict accordingly the future values of rain. The existing statistical models that forecast rain are too complex and also not cost effective. Hence we take the approach of a machine learning model, or to further specify, a deep learning model called Long Short Term Memory (LSTM) to try and predict with some accuracy. This study examines the precipitation patterns over the Lucknow region for a period of 20 years, with the dates ranging from 1st January, 2000 to 31st December, 2019. The accuracy of the LSTM model developed is judged on the basis of Mean Absolute Percentage Error (MAPE), R-square (R2) and Root Mean Squared Error (RMSE) values.近年来，受厄尔尼诺（El Nino）、拉尼娜（La Nina）、全球变暖等环境现象引发的气候变化不确定性影响，降雨预报的重要性日益凸显。印度农业仍高度依赖降雨，在北方邦（Uttar Pradesh）这类地区尤为如此。因此，研发能够分析历史降雨趋势并据此预测未来降雨量的预报系统迫在眉睫。现有降雨预报统计模型不仅过于复杂，且成本效益不佳。为此，我们采用机器学习模型，更具体地说，是长短期记忆网络（Long Short Term Memory，LSTM）这一深度学习模型，以期实现具备一定精度的降雨预测。本研究以勒克瑙（Lucknow）地区2000年1月1日至2019年12月31日共20年的降水数据为研究对象，分析其降水变化模式。所构建的LSTM模型的预测精度，通过平均绝对百分比误差（Mean Absolute Percentage Error，MAPE）、决定系数（R-square，R²）以及均方根误差（Root Mean Squared Error，RMSE）三项指标进行评估。