Meta-learning for water level prediction
- 1Dublin City University, School of Electronic Engineering, Ireland (asma.slaimi2@mail.dcu.ie)
- 2Dublin City University, School of Computing, Ireland
- 4Dublin City University, School of History and Geography, Ireland
- 5Dublin City University, DCU Water Institute, Ireland
- 6Dublin City University, Insight SFI Research Centre for Data Analytics, Ireland
In the Artificial intelligence (AI) sense, meta-learning is the ability of an artificially intelligent machine first to learn how to conduct different complex tasks, taking the principles it utilised to learn one task and applying them to other different tasks. Hence, the general concept of "learning how to learn". Machine learning provides capabilities to learn from past data and generates models for future prediction, which can be helpful for multiple catchment management tasks, such as water elevation monitoring and flood prediction.
Our initial studies focused on predicting and evaluating the ML-based hydrologic time-series models based on their predictive performance. We used eight machine learning algorithms to predict river water levels, including Baseline, Linear, Dense, MultiDense, CNN, RNN, GRU and LSTM techniques. The eight models were employed for one hour ahead of river water level forecasting in 70 hydrometric stations in Ireland. The results show that the NN-based models generally performed well in predicting the water level, with some differences in each model's performance for different stations. These results suggest that a single machine learning model may be sufficient for forecasting river water levels in one location and perform poorly in another. Hence, there is no overall best model; and the selected model may significantly impact the desired results.
This study's main goal was to investigate a meta-learning-based approach for water level prediction. The proposed Meta-learning approach comprises two phases; Learning and meta-learning. The meta-learning process uses the outcomes of the previous experiments to accomplish the Learning Training and Practising phases of the meta-learner. Later the outcome of the previous step will be the Databases to create the learner (learning about learning phase).
Creating meta-learning models can help AI models to generalise learning methods and acquire new skills more quickly. We expect the meta-learning model to adjust well when generalising to previously unknown datasets and environments that have never been encountered during training.
Keywords: Machine learning (ML), meta-learning, water-level prediction, hydrologic time-series forecasting.
How to cite: Slaimi, A., Scriney, M., Hegarty, S., Regan, F., and E. O’Connor, N.: Meta-learning for water level prediction, EGU General Assembly 2023, Vienna, Austria, 24–28 Apr 2023, EGU23-16266, https://doi.org/10.5194/egusphere-egu23-16266, 2023.
