Abstract: Thermal error, the primary cause of machining accuracy loss in high-speed motorized spindle systems, which is essentially caused by the dynamic variation of the spindle's internal thermal state. Current research on spindle thermal characteristics encounters two key problems. Firstly, the core heat-generating components are enclosed in a sealed structure, so the internal thermal state cannot be directly measured by sensors. Secondly, temperature evolution of the motorized spindle under multi-speed conditions presents strong nonlinearity and spatiotemporal coupling, and existing models fail to realize temporal generalization prediction across working conditions. To address this issue, a high-precision temperature prediction model for motorized spindles is proposed to accurately predict internal temperature changes at different rotational speeds. Firstly, a finite element simulation model of spindle temperature field under multiple working conditions is established by simulation software. With full consideration of heat exchange among internal spindle components, temporal temperature data of key internal parts are obtained. Then, a prediction model based on temporal convolutional network (TCN) and gated multi-layer perceptron (gMLP) is constructed, in which the TCN and gMLP modules extract temporal and spatial features respectively. Performance comparisons with mainstream time-series prediction models such as LSTM, GRU, CNN-LSTM, CNN-GRU, and Transformer show that the proposed TCN-gMLP model significantly outperforms the comparative models, with a coefficient of determination (R²) reaching 99.98%, a mean squared error (MSE) of 0.0301, and a root mean squared error (RMSE) of 0.0363.