24AI Application of Data Driven Model in the Forecast of Short-Range Microscale Rainfall to Mitigate the Impact of Climate Change in Tropical Climates

There have been notable changes in the patterns of several climatic elements like atmospheric pressure, temperature, wind, and precipitation. These changes will impact climatic patterns and subsequently affect the continuous circulation of water in the Earth’s hydrosphere. Climate change will affect all regions globally, although the extent of the impact will vary based on factors such as latitude, the level of development, and preparedness for the impending changes.

One of the most evident indicators of climate change is the change in rainfall patterns. Rising temperatures, as reported by the Intergovernmental Panel on Climate Change (IPCC), are projected to significantly impact the distribution and intensity of rainfall, particularly in moisture-convergence climates. These changes are likely to intensify extreme weather events such as storms, heavy rain and floods that lead to the destruction of infrastructure and crops. Also, these shifts can increase the duration and likelihood of drought, leading to famine and wildfires.

Most countries in tropical climates are vulnerable to changes in rainfall. Variations in rain levels can highly impact biodiversity and life in these regions. The changes in rainfall patterns due to climate change further increase the uncertainty in forecasting the occurrence of rainfall, hence have significant negative economic consequences across various sectors, particularly agriculture, telecommunications, and water resource management, with levels of impacts that vary by region and time.

Weather forecasting has advanced over the last half-century, especially in the past two decades, from a resolution of over 750km to as little as 5km. These supercomputers located in operational weather centres run on sophisticated and complex numerical models like the ECMWF and GFS . Even with the advancement in weather forecasting, this success is not resonating in the tropical climate. Forecasting rainfall is crucial in tropical climates. Rain plays a key role in regulating temperature and determining seasons. It exhibits high intensities and is characterised by convectional patterns with significant spatiotemporal variabilities. Despite advancements in weather forecasting, accurate predictions remain challenging in tropical regions due to the localised nature of rainfall, with average cell sizes of less than two kilometres. This high degree of spatial and temporal variability complicates forecasting efforts, as different rainfall patterns can occur even between areas just a few kilometres apart. Therefore, timely and precise rainfall predictions are essential for effective planning and disaster mitigation.

Due to its dense vegetation, rain measuring instruments like radar and satellite systems have poor spatial and temporal resolution. Moreover, the localised nature of rainfall is not well captured due to the lack of local weather measuring systems. The cost of measuring rainfall with sufficiently high resolution for effective disaster prevention and water resource management is exorbitantly high. Additionally, the lack of consistently acquired rainfall data over an extended period covering a large area has hindered the development and effective use of sufficiently accurate prediction models, especially during heavy rainfall and drought periods.

Given that most developing countries have tropical climates, the aim is to design a low-cost rainfall state forecasting system to encourage multi-location deployment due to the small rain-cell size prevalent in this climate. This study focuses on developing and evaluating data-driven short-range rainfall forecasting models for predicting the state of rainfall (Yes- rainfall, No- no rainfall) in a given tropical climate location. Three years of hourly data, from 01/01/2012 to 31/01/2014, were collected from the Alor Setar weather station in the Kedah region of Malaysia. Atmospheric parameters, namely atmospheric pressure, temperature, dew point, humidity, wind direction and wind speed, are used as input parameters. The rainfall depth, initially a continuous variable, was converted to a categorical variable by equating zero values to No (no rain) and all values greater than zero indicative of Yes (rainfall) as the output variable. The forecast is carried out within a given time frame, ranging from one-hour ahead forecasts to twenty-four hours ahead, with intervals of six hours. Thus, five prediction models are designed to forecast the state of rainfall for the next hour, six hours ahead, twelve hours ahead, eighteen hours ahead and twenty-four hours ahead.

The state prediction of rainfall in a tropical climate is achieved using a supervised machine learning (ML) technique, as it is a data-driven modelling approach where the output values of the dataset are known. There are several supervised ML classification techniques, such as artificial neural network (ANN), random forest (RF) support vector machine (SVM) and logistic regression. Among these classification approaches, logistic regression has noteworthy advantages compared to the other methods. It is a probabilistic binary classifier widely used across various fields for prediction and forecasting over the past decades, including forecast applications for weather, precipitation and temperature forecasting. Hence, it is preferred and applied due to the flexibility and simplicity in solving classification and regression problems and the ease of identifying the most important variables used for the model design.

The five models were trained using the first two years as the training dataset, and the test dataset consisted of the third year. The competence of each trained model was evaluated on the logistic regression model assumptions, which include appropriate outcome structure, large sample size, independence of error terms/observation, no influential outliers, linearity of continuous independent variables and the logit, and moderate or absence of multicollinearity. The models were validated using the Hosmer- Lemeshow test to measure their goodness-of-fit. This test is appropriate for logistic regression models with categorical and continuous variables as input variables. Their forecast accuracy are evaluated and compared using the Area Under the Receiver Operating Characteristics curve (AUROC) and the balance accuracy metric. The balance accuracy metric is used instead of accuracy as it takes into account the imbalance in the data, with approximately 4.03% of the response being Yes and 95.97% being No.

The comparative review of the AUROC of each model indicates that the one-hour ahead model has the highest score of approximately eighty percent. Although the accuracy decreases as the forecast window extends, the 24-hour ahead model still performs well. This is likely due to the distinct sinusoidal patterns of tropical rainfall within a 24-hour cycle.