A Short-Term Flood Forecasting Model Using Markov Chain

Flood is one of the disastrous events in Canada, affecting both the sustenance of people and the country’s economy. A short-term flood warning system can be beneficial for small cities as it requires less data and helps different stakeholders to take preventive measures to lessen flood damages. This study developed a first-order Markovian model (MC) as a short-term flood prediction appliance using eighteen years of daily data of two stations along the Fraser River in British Columbia to generate future water level states. Cumulative transition matrix and uniform random numbers help to generate these states on a monthly and seasonal basis. To assess the MC model’s validity, overall prediction percentages of states in terms of correct, overestimation and underestimation percentages and accuracy of predicting individual states are analyzed. The results depict that the monthly MC model predicts future states with an accuracy between 71.3% and 95.2% for both stations. This model can also forecast high flood level (> 5.5 m) states with 78.51% accuracy approximately for both stations but is not suitable for May (Mission station) and June (Hope station) due to large underestimation percentages. The seasonal MC model for ‘Fraser River at Hope’ provides better prediction accuracy (82.83% to 92.44%) in forecasting both flooding and other states. In contrast, the spring and summer MC model of ‘Fraser River at Mission’ station can not predict the higher states due to larger underestimation percentages of 15.38 and 16.45, respectively.