Autonomous Multi-Industry Time Series Prediction Pipeline using Lightweight CNN and Indexed Search on a Local Computer

This study proposes an autonomous time series prediction pipeline that operates in a network-disconnected local environment. The proposed system automatically processes spreadsheet-based user inputs to perform preprocessing, data transformation, prediction, and result generation, and is designed to re-execute the entire pipeline when input changes occur. Multivariate time series data are converted into a multi-channel structure and processed using a lightweight convolutional neural network (CNN); probability-based logits generated by the CNN are accumulated over time and compared with historical vectors to analyze structural similarity and directional changes, enabling multi-step forecasting of up to eight future time points. An index-based search mechanism is further integrated to efficiently compare past states, and prediction results are automatically generated in HTML and CSV report formats for practical use across diverse industrial domains in local environments.