A modular Java-based framework for deploying ONNX time-series forecasting models: a rainfall prediction case study

Bringing deep learning models for time-series forecasting into production has its challenges in the real world, and one of these challenges is transitioning from a Python-based development environment to platform-agnostic software. This paper proposed a modular extensible and easy-to-use Java-based framework for implementing deep learning time-series forecasting models, using simulated rainfall forecasting as a specific case study. One of the benefits of the case study of rainfall forecasting was the number of datasets available to work with and its relevance to environmental monitoring. The framework is able to seamlessly implement Long Short-Term Memory (LSTM) models that were trained in PyTorch and exported in the ONNX (Open Neural Network Exchange) format running inference with DJL (Deep Java Library). Users benefit from the ability to convert Java-native data to DJL compatible tensors via a custom Translator module in real time or batch for prediction. The framework also provides an easy-to-use graphical user interface (GUI) built in JavaFX to allow users to import CSV datasets to predict, visualize results, and export outputs without any advanced programming experience. In the rainfall forecasting case imposed for the case study analysis, the predictive accuracy was limited by the dataset; however, the main purpose of the work was to develop a reusable, accessible, and extensible deployment platform for ONNX-based deep-learning time-series models. The framework provides the foundation to allow for practical use of machine-learning workflows in a variety of applications, including environmental management, logistics, and industrial automation. The modularity of the framework and cross-platform development help to fill the gap with existing deployment technologies which offer a scalable pathway of operationalizing machine learning in practice with modern models such as deep learning. The proposed framework provides accessible solution between advanced model development and deployment covering wide use cases of machine learning.