EL-RMLocNet: A precise deep learning model makes use of a novel k-hop neighbourhood based scheme to capture comprehensive relations of nucleotides in order to generate a rich statistical representation of RNA sequences, LSTM with attention find most relevant features and their dependencies for accurate multi-compartment subcellular lcalization of 4 different RNA classes including mRNA, miRNA, snoRNA, and lncRNA across 2 distinct species including homo sapien and mus musculus. Training and generalizeability of proposed EL-RMLocNet approach are optimized using bunch of neural tricks. A comprehensive performance comparison of EL-RMLocNet approach with state-of-the-art approach shows that EL-RMLocNet approach significantly outperforms previous best performance across a variety of evaluation metrics, achieving top generalizeability and robustness to different sized datasets and subcellular compartment cardinalities. A thorough analysis of potential nucleotides patterns for different RNAs and species make the decision making of proposed approach quite interpretable, indicating a great value for diverse practical bimedical applications. EL-RMLocNet web server enables the user to train and optimize feature extraction and subcellular localization prediction modules of proposed EL-RMLocNet approach from scratch as well as to make inference on the go over new RNA sequences, and download diverse artifacts during the session.