陳佳宏1, *
李奕德2
許志仲3
1國立成功大學地球科學系
2國家太空中心
3國立陽明交通大學智慧系統與應用研究所
*E-mail: koichi@mail.ncku.edu.tw
The current achievements of the project focused on constructing the potential model prediction model, detailing the methodology, design goals, and execution status as of the mid-term review. The model development strategy is structured around three key objectives:ensuring the simultaneous consideration of both temporal and spatial information; redesigning the traditional nxn model architecture to proficiently handle input and output sequence lengths
that may be inconsistent; and expanding the feature channels by integrating multiple inputs, including the F10.7 index, ap index, and sunspot number (SN). The core process for model construction encompasses several stages, beginning with data acquisition, followed by data alignment and data normalization, and culminating in the establishment of training and testing samples. For the predictive system, models such as Metnet 3 and a multi-channel PredRNN have been adopted. The operational configuration involves setting crucial hyperparameters,
such as dynamically adjusted learning rates, employing the Structural Similarity Index Measure (SSIM) as the loss function, and utilizing evaluation metrics like Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and R-squared (R2). Current progress confirms that the model construction phase, incorporating Metnet3 and multi-channel PredRNN, is complete. This implemented model utilizes an input data combination comprising {F10.7/ap/SN/By/Bz/Potential map at t0} to generate an output {Potential map at t1}. Crucially, the lengths of t0 and t1 may be unequal, fulfilling one of the primary construction directions, and extensive testing involving multiple experiments is scheduled to determine the optimal lengths for these sequential inputs and outputs.
關鍵詞：Space weather prediction, potential map, Metnet 3, PredRNN
