Dynamic　multi-objective　optimization　problems　(DMOPs)　have　several　conflicting　and　timevarying　objectives　or　constraints.　To　quickly　follow　the　dynamical　Pareto　optimal　front　(POF)　of　DMOPs,　prediction　model-based　optimization　algorithms　have　been　widely　studied.　However,　in　most　existing　prediction-based　methods,　only　the　linear　relationship　of　historical　solutions　is　studied,　and　complex　correlations　among　the　decision　variables　are　always　ignored.　To　address　this　issue,　the　multi-reservoir　ESN　(MRESN)　based　predictor　is　designed　and　integrated　with　the　multi-objective　evolutionary　algorithm　based　on　decomposition　(MOEA/D),　which　is　called　MRESN-MOEA/D　in　short.　The　comprehensive　relationship　among　the　previous　solutions　is　derived　using　the　MRESN　predictor,　whose　multi-reservoir　structure　projects　the　inputs　into　the　complex　echo-state　space　and　enhances　the　information　sharing　among　the　decision　variables.　To　overcome　the　limitation　caused　by　insufficient　training　solutions,　the　fractal　interpolation　technique　is　implemented　before　MRESN　training.　Then,　the　trained　MRESN　predictor　is　applied　to　produce　the　original　population　for　the　new　environment.　Finally,　MRESN-MOEA/D　is　applied　in　both　simulated　benchmarks　and　an　actual　dynamical　wastewater　treatment　system.　The　experiment　results　illustrate　that　the　proposed　algorithm　outperforms　other　state-of-the-art　methods　with　better　convergence　and　diversity.