In order to address the capacity prediction issue arising from the performance degradation of lithium-ion batteries, this study proposes a BiLSTM-Attention lithium battery capacity prediction model based on Bayesian optimisation with multi-feature fusion. The open-source battery ageing dataset from the University of Maryland is utilised, and the multivariate features are dimensionally simplified by principal component heat map analysis to extract the key multivariate feature elements characterising the capacity degradation. The Bayesian optimisation algorithm's hyper-parameter acquisition process is then employed to construct a bidirectional long- and short-term memory network (BiLSTM-Attention) with an attentional mechanism, which can effectively capture the battery degradation process's temporal dynamic characteristics during battery degradation. The experimental findings demonstrate that the BiLSTM-Attention hybrid model demonstrates remarkable efficacy in battery capacity prediction, with an RMSE of 0.0151 and an R2 of 0.9952.