Desain dan Implementasi Long Short-Term Memory (LSTM) pada Sistem Monitoring dan Data Logger Photovoltaic (PV) untuk Prediksi Daya Keluaran Berdasarkan Fluktuasi Intensitas Cahaya
Fernando Yoga Pratama, Dr. Ir. Atikah Surriani, S.T., M.Eng., IPM.
2026 | Tugas Akhir | D4 Teknologi Rekayasa Instrumentasi dan Kontrol
Fluktuasi intensitas cahaya matahari menyebabkan keluaran daya sistem Photovoltaic (PV) off-grid bersifat dinamis dan sulit diprediksi menggunakan pendekatan konvensional. Penelitian ini mengembangkan sistem monitoring dan prediksi daya berbasis Internet of Things (IoT) menggunakan ESP32, sensor INA219, BH1750, dan DS18B20. Data dikirimkan ke platform ThingsBoard melalui protokol MQTT dengan interval perekaman 15 detik, menghasilkan dataset time-series sebanyak 14.261 sampel selama 11 hari pengamatan. Analisis data menunjukkan adanya fluktuasi daya akibat perubahan kondisi lingkungan serta fenomena PWM cycling pada Solar Charge Controller (SCC) yang memengaruhi karakteristik keluaran sistem. Hasil validasi sensor INA219 menunjukkan rata-rata galat pengukuran tegangan sebesar 0,34?n arus sebesar 1,66%, sehingga data yang diperoleh layak digunakan untuk proses pemodelan. Model prediksi dikembangkan menggunakan arsitektur Deep Long Short-Term Memory (Deep LSTM) multivariat empat lapis dengan masukan berupa intensitas cahaya, suhu panel, tegangan, dan arus pada lookback window 20 timestep. Pada data mentah, model menghasilkan MAPE sebesar 6,75?n R2 sebesar 0,9797. Pada data terfilter yang hanya mencakup kondisi pengisian normal, model menghasilkan MAPE sebesar 6,40%, RMSE sebesar 0,8828 W, MAE sebesar 0,6193 W, dan R2 sebesar 0,9867. Hasil ini menunjukkan bahwa model Deep LSTM multivariat dapat diterapkan untuk prediksi daya keluaran sistem PLTS off-grid skala kecil pada horizon waktu ultra-short-term berbasis data dari sistem monitoring IoT.
Fluctuations in sunlight intensity cause the power output of off-grid photovoltaic (PV) systems to be dynamic and difficult to predict using conventional approaches. This research develops an Internet of Things (IoT)-based power monitoring and prediction system using the ESP32, INA219, BH1750, and DS18B20 sensors. Data is sent to the ThingsBoard platform via the MQTT protocol with a 15-second recording interval, resulting in a time-series dataset of 14,261 samples over 11 days of observation. Data analysis shows power fluctuations due to changes in environmental conditions and the PWM cycling phenomenon in the Solar Charge Controller (SCC), which affects the system’s output characteristics. The validation results for the INA219 sensor showed an average measurement error of 0.34% for voltage and 1.66% for current, making the data suitable for use in the modeling process. The prediction model was developed using a four-layer multivariate Deep Long Short-Term Memory (Deep LSTM) architecture with inputs of light intensity, panel temperature, voltage, and current at a lookback window of 20 time steps. On raw data, the model produced a MAPE of 6.75% and an R2 of 0.9797.On filtered data, which only included normal charging conditions, the model produced a MAPE of 6.40%, an RMSE of 0.8828 W, an MAE of 0.6193 W, and an R2 of 0.9867. These results demonstrate that the multivariate Deep LSTM model can be applied to predict the output power of small-scale off-grid solar power plants (PLTS) over an ultra-short-term time horizon based on data from an IoT monitoring system
Kata Kunci : Photovoltaic (PV), Deep LSTM, Internet of Things (IoT), Prediksi Daya, Intensitas Cahaya Matahari, Data Logger, Ultra-Short-Term Forecasting.