The deployment of deep learning models directly onboard satellites is emerging as a promising approach to real-time space-weather monitoring. However, applying this paradigm to solar flare forecasting introduces three challenges: limited computational resources, the susceptibility of onboard weight storage to radiation-induced multiple-bit upsets (MBUs), and variations in power availability across orbital phases. In this work, we propose N-Slim, a network slimming framework that addresses all three challenges through a novel N-balanced structured channel pruning method. N-Slim proceeds through three phases: baseline pretraining, sparsity learning, and iterative slimming. The core idea of N-balanced pruning is to partition each layer’s output channels into non-overlapping groups of width N and distribute retained channels approximately uniformly across groups, preventing retained channels from clustering together and limiting the number of active channels that a single MBU event can simultaneously corrupt. By applying this method across a geometric sequence of group sizes, N-Slim generates a portfolio of subnets spanning a range of channel pruning rates and inference costs, providing a set of candidate models that could, in principle, support power-adaptive inference during different orbital phases. N-Slim was evaluated on VGG, ResNet, and DenseNet using solar active-region magnetogram images from the Solar Dynamics Observatory (SDO). Experiments demonstrate that N-Slim achieves competitive or improved forecasting accuracy across a range of channel pruning rates, with substantial MBU robustness gains on ResNet and DenseNet and comparable robustness to standard pruning on VGG.