This paper introduces a Deep Reinforcement Learning (DRL) framework for long-short portfolio optimization in China's A-share market. The key innovation is the integration of short-selling mechanisms that adhere to real-world trading rules, enabling low-risk arbitrage through hedging correlated assets. The framework uses deep neural networks to process multi-dimensional financial time series and employs the mean Sharpe ratio as the reward function to optimize risk-adjusted returns. Empirical results demonstrate that the DRL model with short-selling significantly outperforms traditional long-only strategies, achieving consistent positive returns, superior risk-adjusted performance, and reduced maximum drawdown. The model also exhibits a robust investment style by strategically avoiding underperforming assets and maintaining balanced capital allocation. This work contributes to portfolio theory and provides practical methodologies for quantitative investment.