Feature enhanced multistream RNN for growth phase prediction of Euglena sp. microalgae in an IoT-based outdoor cultivation environment

Accurately predicting growth phases in microalgae cultivation is crucial for optimizing biomass production. IoT systems provide convenience in monitoring the cultivation environment in real time. However, the specific challenge of predicting microalgae growth phases still needs to be effectively addressed using IoT-based sequential monitoring data. This study introduces a novel architecture, the Feature-Enhanced Multistream Recurrent Neural Network (FEM-RNN), integrated with an IoT microalgae monitoring system to predict the growth phase of cultured microalgae, especially Euglena sp. species. The proposed method utilizes a dual channel architecture of recurrent neural networks to assess temporal environmental data, i.e., turbidity, temperature, and light intensity, acquired by the IoT system. One channel leverages all input features, while the other is specified for turbidity data. The proposed model was evaluated using a primary dataset collected by an IoT monitoring system from microalgae cultivation in outdoor environments. Three versions of FEM-RNN, i.e., utilizing the base model of Vanilla RNN, LSTM, and GRU, have been assessed in various sizes of window data. The performance of the FEM-RNN models increases with expanding the window size. All the variant models demonstrate high performance at window size 60, and the LSTM-based FEM-RNN demonstrates outstanding performance and stability beginning at the window size. At the size of the window, model performance has been compared to the traditional model, namely Vanilla RNN, LSTM, and GRU models, as well as CNN, Transformer, and SVM. The results show that the proposed model outperforms the conventional models, with an accuracy of 0.978, 0.989, and 0.951 for FEM-RNN based on Vanilla RNN, LSTM, and GRU, respectively. The results indicate that the FEM-RNN effectively predicts the microalgae growth phase utilizing IoT-based monitoring data. © 2025 The Authors