To address critical challenges in long-sequence energy consumption prediction for electric vehicles (EVs), including memory decay, high computational complexity of attention mechanisms, and insufficient adaptability to dynamic driving conditions, this paper proposes a memoryaugmented attentionbased prediction model termed deep Q networkadaptive memory augmented gating (DQNAMAG). The model is built upon an adaptive memoryaugmented gating (AMAG) network that incorporates a threelevel collaborative memory architecture consisting of shortterm memory, longterm neural memory, and persistent memory. A surprisedriven decay mechanism is introduced to enhance the modeling of battery degradation and abrupt operating condition variations, enabling effective capture of multiscale temporal dependencies in longhorizon energy consumption sequences. Furthermore, an adaptive Nystrm attention (ANSA) mechanism is developed to perform lowrank approximation of the attention matrix via the Nystrm method with adaptive sampling dimension adjustment. This reduces the computational complexity from O(T2) to O(T·r), significantly improving efficiency and realtime performance in longsequence scenarios. An adaptive multiscale spatiotemporal attention mechanism (AMSTA) mechanism and a hypernetworkbased dynamic forward model are additionally introduced to enhance deep crossmodal fusion between road condition images and battery management system (BMS) timeseries data, strengthening environmental perception capability. Moreover, the AMAG network is embedded into a reinforcement learning framework, where temporal difference learning provides temporalconsistency regularization and enables selfcalibrated parameter optimization. Experimental results based on five years of realvehicle operational data from two vehicle types demonstrate that the proposed model achieves a mean absolute error (MAE) below 02%, a root mean square error (RMSE) below 03%, and Rsquared (R2) above 995% under different stateofhealth (SOH) conditions. The model exhibits superior stability and generalization performance in long-sequence prediction and battery degradation scenarios, significantly outperforming mainstream models such as Transformer, Informer, Mamba, and LSTM.