Accelerated edge devices, like Nvidia's Jetson with 1000+ CUDA cores, are increasingly used for DNN training and federated learning, rather than just for inferencing workloads. A unique feature of these compact devices is their fine-grained control over CPU, GPU, memory frequencies, and active CPU cores, which can limit their power envelope in a constrained setting while throttling the compute performance. Given this vast 10k+ parameter space, selecting a power mode for dynamically arriving training workloads to exploit power-performance trade-offs requires costly profiling for each new workload, or is done \textit{ad hoc}. We propose \textit{PowerTrain}, a transfer-learning approach to accurately predict the power and time consumed when training a given DNN workload (model + dataset) using any specified power mode (CPU/GPU/memory frequencies, core-count). It requires a one-time offline profiling of $1000$s of power modes for a reference DNN workload on a single Jetson device (Orin AGX) to build Neural Network (NN) based prediction models for time and power. These NN models are subsequently transferred (retrained) for a new DNN workload, or even a different Jetson device, with minimal additional profiling of just $50$ power modes to make accurate time and power predictions. These are then used to rapidly construct the Pareto front and select the optimal power mode for the new workload. PowerTrain's predictions are robust to new workloads, exhibiting a low MAPE of $<6\%$ for power and $<15\%$ for time on six new training workloads for up to $4400$ power modes, when transferred from a ResNet reference workload on Orin AGX. It is also resilient when transferred to two entirely new Jetson devices with prediction errors of $<14.5\%$ and $<11\%$. These outperform baseline predictions by more than $10\%$ and baseline optimizations by up to $45\%$ on time and $88\%$ on power.