Keywords: Deep Learning, Time Series Prediction, LSTM Model, Hidden Layers, Time Scale, Feature Selection, Dataset Split Ratio, Normalization Methods, Model Optimization, Automated Training, Batch Parameter Tuning, Prediction Accuracy.
This project entails the code implementation for the Time Series Prediction Project, which is the second part of the Spatio-Temporal Big Data Analysis and Processing course internship.
In this internship, under the guidance of my professor and senior colleagues, I delved into the scientific questions and basic methods of deep learning and time series prediction, as well as the fundamental principles of the LSTM model and factors influencing its performance. I conducted in-depth experimental research on the impact of different hidden layers, time scales, feature selections, dataset split ratios, and normalization methods on the prediction results of the LSTM model.
Additionally, I refactored and optimized the reference project, dividing the original .ipynb file into five key modules, thus enabling a richer array of experiments and facilitating modification of specific details. The refactored project exhibits low code coupling, supports batch experiments, and can automatically save experiment parameters and visualize prediction results. In the internship project, I designed an automated training mode, namely batch parameter tuning by traversing the parameter list, which to some extent lightens the experimental burden on researchers. The experimental results of this internship indicate that richer features beneficial to prediction, finer time scales, and Standardization normalization methods all contribute to enhancing the prediction accuracy of the LSTM model.