Forecasting Water Levels in Twente Canal Using Time Series Analysis

Introduction: Water levels in the Twente canal have experienced historical lows in year 2018, which affected the supply chain and the logistic operations of the businesses that use the canal to be hindered. As a result, it has become apparent that predicting these unprecedented changes in water levels is crucial to prevent such negative effects again. This research aims to provide an answer to how forecast models for the Twente canal could be made using time series analysis. The research is part of a bigger project in which a digital twin of the Twente canal is being made. The forecasts provided by this study serve the purpose of monitoring the water levels in this project and also provide a basis for any further models to be developed for the project. Theoretical Framework: As the relevant modelling techniques had to be identified as well as a theory needed for the research to be based upon, a theoretical framework was made utilising the existing literature. Several models such as ARMA, ARIMA, SARMA, SARIMA, PARMA, ARIMAX, SARIMAX and MLR were identified as a result of this research. Additionally, several important properties such as seasonality and stationarity were described. Data Understanding and Transformation: Data used in this report were acquired from Rijkswaterstaat and KNMI, both which are public sources of data. In general, data was of high quality and did not require much cleaning in terms of outliers and measurement types. However, the data gathered from Rijkswaterstaat was measured every 10 minute interval, which had to be transformed into a daily average value. After the data was cleaned and transformed into a daily average value, properties of the dataset were examined. Data exhibited low variance and standard deviation which could be the result of averaging the values for daily measurements. Additionally, data was found to be normally distributed, non-stationary and non-seasonal. Finally, several exogenous variables, for which the data was gathered from KNMI, were investigated for use in modelling. Unfortunately, none of there were deemed suitable for different reasons. Modelling: After the data was explored and transformed, it was ready to be modelled. Several possible models depending on the properties highlighted before were identified. Mainly ARIMA models were found suitable for modelling time series. Additionally, although expected to not add value, an ARIMAX model using temperature as an exogenous variable was modelled for research purposes. According to the theoretical framework and methods proposed in it, parameters for the models were estimated. After the initial estimations, models were fitted and compared on their information criterion. Based on this comparison, models with the lowest criterions were chosen to be actually modelled for forecasting. Results and Conclusion: In general, there are mixed results from the modelling phase. In particular, long term forecasts were a failure due to the predicted values converging to the sample mean of the training dataset. Several reasons as to why this behaviour occurs could be unincorporated seasonality and low variance in the dataset. On the other hand, short term predictions were highly accurate and were able to showcase the patterns that the actual values follow. Unfortunately, it is arguable how these short term forecasts could be utilised.