Artificial neural network: physico-chemical and macronutrients parameters in an aquaponic system / Qistina Khadijah Abd Rahman, T.s Mohamed Syazwan Osman and Dr Samsul Setumin

Aquaponic system which integrates conventional aquaculture and hydroponic in one closed-loop system plays a significant role as an alternative way to produce very least waste effluent to the environment by recycling back the nutrients (fish waste) for plant growth. Prediction of water quality parameter in wastewater using conventional mathematical modeling is very complex to simulate and model out the system. Therefore, this paper proposed ANN model to evaluate graph comparison between the performances of the actual data from aquaponics activity and forecast data from simulated artificial neural network (ANN). Then, the best algorithms will be selected in a variety of neuron numbers of the ANN’s model. The parameter such as pH, DO, TAN, and percent total sludge of Phosphorus (P) and Nitrogen (N) were investigated by taking the input and target data value from the selected research paper covering the fields of aquaponic. In this study, Levenberg-Marquardt (LM) and Scaled Conjugate Gradient (SCG) training function were used to measure those parameters to obtain the predict values. For parameter pH, DO, TAN, ranges hidden neurons of 4, 6, 8, 10, 12, 13 neurons were studied. Meanwhile, ranges hidden neurons of 3, 4, 6, 9, 12 neurons were studied for total sludge (P and N). Different range neurons value was used for pH, DO, TAN, and Total Sludge (P and N) due to different input data found in the literature. The outputs from the model of training function LM show the most optimum neuron number for each parameter of pH, DO, TAN at neuron 6. As for total sludge (N and P), the most optimum neuron number at neuron 3. For the training function SCG, the most optimum neuron number at neuron 4 for each parameter pH, DO, TAN and at neuron 9 and neuron 4 were the most optimum neuron number for parameter Total Sludge (N and P). The result for the most optimum neuron number can be explained by the value of Sum Squared Error (SSE) and Mean Absolute Percentage Error (MAPE%) with the lowest value. The investigated forecast parameters of the trained neural network according to correlation coefficient (R) and Mean Square Error (MSE) showed LM performed better rather than SCG.