Causal　discovery　from　river　runoff　data　aids　flood　prevention　and　mitigation　strategies,　garnering　attention　in　climate　and　earth　science.　However,　most　climate　causal　discovery　methods　rely　on　conditional　independence　approaches,　overlooking　the　non-stationary　characteristics　of　river　runoff　data　and　leading　to　poor　performance.　In　this　paper,　we　propose　a　river　runoff　causal　discovery　method　based　on　deep　reinforcement　learning,　called　RCD-DRL,　to　effectively　learn　causal　relationships　from　non-stationary　river　runoff　time　series　data.　The　proposed　method　utilizes　an　actor-critic　framework,　which　consists　of　three　main　modules:　an　actor　module,　a　critic　module,　and　a　reward　module.　In　detail,　RCD-DRL　first　employs　the　actor　module　within　the　encoder-decoder　architecture　to　learn　latent　features　from　raw　river　runoff　data,　enabling　the　model　to　quickly　adapt　to　non-stationary　data　distributions　and　generating　a　causality　matrix　at　different　stations.　Subsequently,　the　critic　network　with　two　fully　connected　layers　is　designed　to　estimate　the　value　of　the　current　encoded　features.　Finally,　the　reward　module,　based　on　the　Bayesian　information　criterion　(BIC),　is　used　to　calculate　the　reward　corresponding　to　the　currently　generated　causal　matrix.　Experimental　results　obtained　on　both　synthetic　and　real　datasets　demonstrate　the　superior　performance　of　the　proposed　method　over　the　state-of-the-art　methods.