In　wastewater　treatment　plants　(WWTPs),　the　prediction　of　effluent　ammonia　nitrogen　(NH4-N)　concentration　is　vital,　which　is　a　major　cause　of　lake　eutrophication.　To　solve　this　problem,　the　evolving　deep　delay　echo　state　network　(EDDESN)　is　proposed.　Firstly,　the　EDDESN　is　decomposed　into　several　serially　connected　sub-reservoirs,　which　are　inserted　delay　units　to　learn　the　temporal　relationships　within　sequence　data.　Secondly,　the　input　and　reservoir　internal　weights　are　generated　by　a　singular　value　decomposition-based　matrix　design　strategy,　which　can　reduce　searching　dimensions　and　guarantee　the　echo　state　property　(ESP).　Moreover,　the　architecture　hyperparameters　and　weights　of　EDDESN　are　simultaneously　optimized　by　a　competitive　swarm　optimizer　(CSO)-based　two-stage　optimization　approach.　Finally,　the　experimental　results　on　practical　NH4-N　dataset　and　simulated　Mackey-Glass　time　series　demonstrate　the　superiority　of　EDDESN　as　compared　with　other　time　series　prediction　approaches.