Location: Crop Production Systems Research
Project Number: 6066-22000-059-08-S
Project Type: Non-Assistance Cooperative Agreement
Start Date: Sep 1, 2009
End Date: Aug 31, 2014
To develop new methods based on wavelet analysis, artificial neural networks and heteroscedasticity time series analysis for parameter forecasting in crop management.
A new method has been proposed to improve performance in forecasting of non-linear, non-stationary time series over the conventional Box-Jenkins method. This method integrates wavelet transforms and artificial neural networks, providing the ability of function approximation for optimal time series pattern recognition. In order to model data sets that depict volatility, heteroscedasticity time series models have been established to outperform the Box-Jenkins method. Many parameters are important in crop management, such as soil temperature, soil moisture, solar radiation, wind run, pan evaporation, rainfall, relative humidity and air temperature. Evaluation of these parameter time series indicated the features of non-linearity, non-stationary, and some degree of volatility. Therefore, to provide reliable and robust forecasting of the parameters, development of the new time series analysis methods is necessary instead of using the Box-Jenkins method, which performs poorly under these conditions. The new methods will be developed based on existing prototypes mutually designed by USDA ARS and Texas A&M University for robust time series analysis and forecasting of dynamic systems. USDA ARS will provide recorded data from his experimental fields as soil temperature, soil moisture and solar radiation. USDA ARS has supplied locally obtained and processed weather data for initial system development and will supply weather, lysimeter, and pan evaporation data as required for system refinement and verification. System prototypes were originally proposed for industrial and social economic system analysis. In the design of the new systems, the Box-Jenkins method is used as a benchmark. The wavelet-neural network method and the heteroscedasticity time series models are compared with the conventional method to demonstrate the advantages in model specification, data characterization, and forecasting ability.