Sistem pressure tank pada produksi Bio-CNG (Compressed Natural Gas berbasis biomassa) data bersifat dinamis dan nonlinier, sehingga pengendalian tekanan presisi dengan metode konvensional menantang. Penelitian ini membandingkan dua pemodel prediksi tekanan satu langkah ke depan berbasis deret waktu Artificial Neural Network (ANN) dengan Simple recurrent dan Long Short-term Memory (LSTM) untuk memodelkan keterkaitan antara duty cycle Pulse Width Modulation (PWM) yang 0–255 dipetakan ke 0–100% dan tekanan tangki (0–1,2 bar). Data diproses melalui resampling sebesar 10 Hz, windowing 5 detik, pelabelan, dan normalisasi berbasis waktu guna mencegah leakage, pelatihan dan pengujian dilakukan pada split per level PWM (0, 55, 100, 150, 200, 255) serta uji perangkat nyata (open-loop) dengan PWM piecewise-constant berganti setiap 10 s hingga 1,2 bar. Pada uji lapangan open-loop, LSTM mencapai MAE (Mean Absolute Error) sebesar 0,014 bar, RMSE (Root Mean Squared Error) sebesar 0,024 bar, R² (Koefisien Determinasi) sebesar 0,991 (≈1,17% dan 2,00% dari full-scale 1,2 bar), sedangkan ANN dan simple recurrent meraih MAE sebesar 0,018 bar, RMSE sebesar 0,026 bar, R² sebesar 0,984, selama pengujian, LSTM mempertahankan RMSE/MAE lebih rendah dan R² sangat tinggi, sementara ANN cenderung memuluskan ekstrem sehingga amplitudo puncak tereduksi. Integrasi model sebagai koreksi feed-forward pada PID (NN-PID, set-point 0,80) menunjukkan LSTM-PID mencapai T₉₀ ≈ 32–34 s, waktu steady state (±5%) ≈ 36–38 s, overshoot ≈ 2–3%, dan deviasi steady state ≈ ±0,01 bar, kinerja ini melampaui PID only (T₉₀ ≈ 41–42 s, stabil di akhir rekaman) dan ANN-PID (T₉₀ ≈ 40–45 s, steady state ≈ 50 ± 5 s). Berdasarkan metrik dan respons transien tersebut, LSTM dipilih sebagai modul prediktif dasar, dan konfigurasi LSTM-PID direkomendasikan sebagai rancangan kontrol akhir untuk prediksi pada sistem Bio-CNG.
The pressure-tank in Bio-CNG production is dynamic and nonlinear, making precise pressure control challenging with conventional methods. This study compares two one-step-ahead time-series predictors—an artificial neural network (ANN) with a simple recurrent feature and a long short-term memory (LSTM) to model the mapping between pulse-width-modulation (PWM) duty cycle (0–255 mapped to 0–100%) and tank pressure (0–1.2 bar). Data were processed via 10 Hz resampling, 5 s windowing, labeling, and time-based normalization to prevent leakage. Models were trained/tested on splits per PWM level (0, 55, 100, 150, 200, 255) and validated on a real open-loop test using a piecewise-constant (zero-order hold) PWM that changes every 10 s until 1.2 bar. In field tests, LSTM achieved MAE (Mean Absolute Error) 0.014, RMSE (Root Mean Squared Error) 0.024, R² (Coefficient Determination) 0.991 (≈1.17% and 2.00% of 1.2 bar full-scale), outperforming ANN and simple recurrent (MAE 0.018, RMSE 0.026, R² 0.984). Perlevel analysis shows LSTM maintains lower RMSE/MAE and very high R², while ANN tends to smooth extremes, reducing peak amplitudes. When integrated as a predictive feed-forward with PID (NN-PID, 10 Hz, set-point 0.80), LSTM-PID reaches T₉₀ ≈ 32–34 s, settling time (±5%) ≈ 36–38 s, overshoot ≈ 2–3%, and steady deviation ≈ ±0.01 surpassing PID-only (T₉₀ ≈ 41–42 s, stabilizing near test end) and ANN-PID (T₉₀ ≈ 40–45 s, settling ≈ 50 ± 5 s). Overall, LSTM is selected as the predictive core, and the LSTM-PID configuration is recommended as the final realtime control design for look-ahead pressure regulation in Bio-CNG systems.
