“Machine learning based Intelligent weather monitoring and prediction system”

Abstract

The problem of weather prediction for agricultural domain is of prime importance for the agriculture experts, farmers and the research institutions across Ethiopia. This research proffered time series and machine learning modeling techniques for the design, development and implementation of light weight, easy to deploy models for the meteorology centers and researchers in the field of weather forecasting for Ethiopia. Totally five important weather parameters named TEMPERATURE, PRECIPITATION, SUNSHINE HOURS, RELATIVE HUMIDITY and RAINFALL were selected for this research in Adama region. The data was collected from 44 meteorology stations in the region. Past ten years data for 33 variables was obtained from Adama Meteorology center and Addis Ababa meteorology center. Before the machine learning models were actually proposed, a thorough study of existing models for the weather prediction task was performed by the research team. Major limitation of existing, traditional models used by Adama meteorology center and nearby weather stations was the requirement of high performance computing resources, which is not available in every center in Ethiopia. These models need continuous sequence of observations for doing model correction and perturbations. High Performance Computing environment need highly trained staff and continuous training and development activity for the task of weather modeling and prediction, which is difficult to provide in case of rural areas of Ethiopia. In this research, machine learning based weather prediction models were proposed as an alternative way of doing this task. Unlike traditional models which are based on grid based finite element method, and simulate the weather phenomenon across the large (full) span of the geographical area, Machine learning models work on the principle of learning the patterns in the observed data from the recent past. The problem of weather forecasting was formulated in terms of uni-variate and multivariate time series models which were ensembled across yearly observation data to obtain greater confidence in the prediction results. The Block Bootstrap sampling technique5 with averaging based ensemble of base models was proposed. Different models were developed to predict five parameters in short, medium and long term settings. The base models used in this research included moving average model, exponential smoothing model, auto-regressive models, autorgeressive neural network model and vector autoregression model. On experiments, for the short term predictions up to 15-20 days, uni-variate ensemble models were found to be more effective, particularly Exponential Smoothing Model and Auto Regressive Integrated Moving Average (ARIMA) model were able to learn the past behavior of selected weather parameters with least mean absolute error. The mean absolute error reported by the Exponential Smoothing model and Auto-regressive integrated models on the short term weather prediction was 0.97 and 0.93 for maximum temperature. Similarly above two models have shown better performance for the remaining weather parameters in short term as well as medium term case, which suggest that the proposed models can be used in weather prediction and creating advisories for agriculture sector as agriculture is more sensitive towards short and medium term changes in weather parameters. Also, auto-regressive multivariate form of neural network was found to be useful and shown a forecast accuracy of 94% which is comparable to existing models like Weather Research and Forecasting (WRF). This research has provided a direction for the Ethiopian meteorology centers to adopt the machine learning in weather prediction tasks as compared with the traditional models. The comparative results and accuracy in the prediction of short term and medium term weather forecasts with less number of resources and simple script based execution of prediction task have encouraged meteorology personal to learn and use techniques proposed under this research. Introducing machine learning models in the prediction task was a new topic for meteorology center in Adama, as the previous system was based on manual data processing and traditional software like Leap and Excel sheet were used for regression based predictions. The current research have successfully introduced and involved the professionals from Adama meteorology center in the process of model6 building, validation and testing. In the future, meteorology centers across Ethiopia have a plan to implement High Performance Computing environment, which is essential for the implementation of traditional finite elements based weather models as well as distributed machine learning models.