End-to-end SOM–AE–LSTM pipeline. Solar disk observations (continuum image, unsigned magnetogram, and μ map) are clustered pixel-wise by the SOM into a 400-node cluster-index map, compressed by the convolutional AE into a 32-dimensional feature vector z_t via global-average pooling, concatenated with the contemporaneous TSI measurement, and processed by the LSTM to produce multihorizon TSI forecasts. Training proceeds sequentially through the three stages; all parameters are estimated from training-period data only.