Foretell: Forecasting environmental data through enhanced lstm and l1 regularization

Abstract

Big data analytics progressively takes over the universe, and the prediction tends to take advantage of Big data analytics to yield the incredible results. The deep learning algorithms procure higher priority than the machine learning in the prediction systems. The traditional weather prediction exploits the observations of prevailing atmospheric conditions that accumulate simultaneously either from the trained observers or the numerical prediction model. However, the weather forecasting is an arduous task due to the dynamic and uncertainty of data. The proposed system plans to use the neural network for weather forecasting to overcome these shortcomings. Recently, the Long-Short Term Memory (LSTM) network based weather forecasting has gained popularity in machine learning. It significantly reflects the superior ability to model the temporal data sequences along with the long-term dependency through the memory blocks. However, on account of memory blocks along with the loop structure leads to over fitting issues. In order to tackle this issue, this work presents FORcasting Environmental data Through Enhanced LSTM and L1 regularization (FORETELL) that extends the existing LSTM model with the two methods such as optimal neuron selection method and the regularization method. The optimal neuron selection method constitutes the FORETELL system, as it can learn the complex data sequences without longer period and overfit of data. Instead of processing the entire vast feature of the data sequence, the regularization method captures the potential features to avert the overfitting constraint that ensures the noise-free system with more excellent performance. Conclusively, the FORETELL is evaluated using the weather dataset to demonstrate the superior performance than the existing Sequence to Sequence Weather Forecasting (SSWF) method.