Abstract:

Forest fire is not easily handled by traditional analysis methods and steady-state neural network because it is a dynamic process over space. A spatial-temporal integrated forecast method of forest fire was proposed in this paper by combining dynamic recurrent neural network(DRNN) and autoregressive integrated moving average(ARIMA) model. The approach first forecasts time series by ARIMA model, and reveals the hidden spatial correlations among forest fire data by DRNN， and then combines the spatial and temporal forecast results based on statistic regressions to produce the final spatial-temporal integrated forecast result. The principle and modeling procedure of the model were illustrated with a case study of forest fire area forecast in Guangdong， and then the forecast accuracy was validated. The results showed that the forest fire area could be forecasted exactly and effectively by the spatial-temporal integrated forecast model because the spatial correlations among data were taken into consideration. Compared with the pure ARIMA model， the forecast precision of the model was apparently improved. The integrated model was also proved to be good efficient in forecasting dynamic change of events over space such as a forest fire.

Key words: dynamic recurrent neural network, ARIMA model, forest fire, spatial-temporal integrated forecast