Correo Electrónico
DNINFOA - SIA
Bibliotecas
Convocatorias
Identidad U.N.
Publicado
Densely Connected Neural Network for Predicting the Pres- ence of Photovoltaic Power Generation in Electric Utility Customers.
Red neuronal densamente conectada para predecir existencia de generación de energía electrica fotovoltaica en clientes de distribuidoras eléctrica
DOI:
https://doi.org/10.15446/sicel.v12.121230Palabras clave:
clasification, prossumer, solar radiation, neuronal network (en)Perfiles de carga, Redes Neuronales, Clientes de Distribuidoras electricas, Deep Learning, Clasificacion de prosumidores, Radiación solar (es)
Descargas
This work explores the use of a densely connected neural network to identify the presence of photovoltaic (PV) power generation in customers of electric utility cooperatives, based solely on their daily electricity demand profiles and local solar radiation data. The core idea is to detect characteristic patterns in energy consumption that suggest behind-the-meter solar generation—such as drops in measured demand during peak sunlight hours.
The dataset was built from approximately 18 months of hourly measurements collected from five utility cooperatives. Due to confiden- tiality agreements, the cooperatives are not named. Only one of them is known to have PV installations, which allowed us to label the data for training and testing purposes. Each data point includes 24 hourly demand values, 24 hourly solar radiation values, and a binary label indicating whether the customer is a prosumer (i.e., has PV generation).
To prepare the data, we reshaped the records into a format suitable for neural network input, cleaned missing or inconsistent entries, and applied standardization to improve training efficiency. The final model architecture consists of an input layer with 48 neurons, followed by two hidden layers with 1000 and 2000 neurons, respectively. All layers use sigmoid activation functions. The output layer contains a single neuron, with a sigmoid activation to predict the probability of PV presence. The model was trained using the Adam optimizer, binary cross-entropy loss, and accuracy as the main performance metric.
We trained the model over 10 epochs, observing that performance on the validation set plateaued after six epochs—this early stop- ping point was selected to avoid overfitting. Results show that the model is able to distinguish between prosumer and non-prosumer profiles with promising accuracy, highlighting its potential for utilities to detect unregistered PV systems and support smarter grid plan- ning.
Este trabajo explora el uso de una red neuronal densamente conectada para identificar la presencia de generación fotovoltaica (FV) en clientes de cooperativas eléctricas, utilizando únicamente los perfiles diarios de demanda eléctrica y datos locales de radia- ción solar. La idea central es detectar patrones característicos en el consumo de energía que sugieran generación distribuida detrás del medidor, como caídas en la demanda medida durante las horas de mayor radiación solar.
El conjunto de datos fue construido a partir de aproximadamente 18 meses de mediciones horarias tomadas en cinco cooperativas eléctricas. Por motivos de confidencialidad, no se revelan los nombres de las entidades. Solo una de ellas se sabe que posee instala- ciones fotovoltaicas, lo que permitió etiquetar los datos para su uso en entrenamiento y prueba. Cada registro incluye 24 valores horarios de demanda, 24 de radiación solar y una etiqueta binaria que indica si el cliente es prosumidor (es decir, si genera energía solar).
Para preparar los datos, se reorganizaron en un formato compatible con redes neuronales, se eliminaron valores faltantes o inconsis- tentes, y se aplicó una estandarización para mejorar la eficiencia del entrenamiento. La arquitectura del modelo incluye una capa de entrada con 48 neuronas, seguida de dos capas ocultas de 1000 y 2000 neuronas respectivamente, todas con funciones de activación sigmoide. La capa de salida consiste en una sola neurona con activación sigmoide que predice la probabilidad de pre- sencia fotovoltaica. El modelo fue entrenado con el optimizador Adam, función de pérdida binary cross-entropy, y precisión como métrica principal.
El entrenamiento se realizó durante 10 épocas, aunque se observó que el rendimiento en el conjunto de validación se estabilizaba alrededor de la sexta, por lo que se optó por detener el proceso en ese punto para evitar sobreajuste. Los resultados muestran que el modelo logra distinguir con precisión perfiles de prosumidores y no prosumidores, demostrando su potencial para ayudar a las distribuidoras a detectar instalaciones fotovoltaicas no registradas y mejorar la planificación de la red.
Descargas
Citas
[1] F. B. Y. Yuan, K. Dehghanpour, Z. Wang, y Y. Guo, "Disaggre-gating customer-level behind-the-meter PV generation using smart meter data and solar exemplars," IEEE Trans. Power Syst., vol. 36, no. 4, pp. 3456–3466, 2021.
[2] D. L. Donaldson, "Effective solar prosumer identification using net smart meter data," Int. J. Electr. Power Energy Syst., vol. 118, 105823, 2020. [Online]. Available: https://doi.org/10.1016/j.ijepes.2020.105823.
[3] A. F. Moreno Jaramillo, J. Lopez-Lorente, D. M. Laverty, J. Martinez-del-Rincon, y A. M. Foley, "Effective identification of distributed energy resources using smart meter net-demand data," IET Smart Grid, vol. 5, no. 2, pp. 120–135, Mar. 2022. [Online]. Available: https://doi.org/10.1049/stg2.12056.
[4] A. Zaboli, "A comprehensive review of behind-the-meter distributed energy resources (DERs): Forecasting methods, models, and applications," Energies, vol. 17, no. 11, art. no. 2534, Jun. 2024. [Online]. Available: https://doi.org/10.3390/en17112534.
[5] D. P. Kingma y J. Ba, "Adam: A method for stochastic opti-mization," Proc. 3rd Int. Conf. Learn. Representations (ICLR), San Diego, CA, USA, 2015. [Online]. Available: https://arxiv.org/abs/1412.6980
[6] I. Goodfellow, Y. Bengio, y A. Courville, Deep Learning. Cambridge, MA, USA: MIT Press, 2016.
[7] M. Sokolova y G. Lapalme, "A systematic analysis of perfor-mance measures for classification tasks," Inf. Process. Manage., vol. 45, no. 4, pp. 427–437, Jul. 2009. [Online]. Available: https://doi.org/10.1016/j.ipm.2009.03.002.
[8] Y. LeCun, Y. Bengio, y G. Hinton, "Deep learning," Nature, vol. 521, no. 7553, pp. 436–444, May 2015. [Online]. Available: https://www.nature.com/articles/nature14539.
[9] T. Chen et al., "Electric load forecasting using deep learn-ing: A review," Energy Rep., vol. 6, pp. 1747–1767, 2020, doi: 10.1016/j.egyr.2020.09.033.
[10] R. M. D. Souza, L. N. Ribeiro, y P. H. M. de Lima, "Deep learn-ing-based photovoltaic power prediction," Renew. Energy, vol. 167, pp. 118–130, 2021, doi: 10.1016/j.renene.2020.11.092.
[11] F. Chollet, Deep Learning with Python, 2nd ed. Manning Publications, 2021.
[12] J. Brownlee, Deep Learning for Time Series Forecasting. Machine Learning Mastery, 2018.
[13] A. M. Foley et al., "Battery storage and demand response for grid balancing: A review," Renew. Sustain. Energy Rev., vol. 62, pp. 1230–1242, 2016, doi: 10.1016/j.rser.2016.05.022.
[14] H. Lund et al., "Energy storage and smart energy systems," Int. J. Sustain. Energy Plann. Manage., vol. 11, pp. 3–14, 2016.
[15] L. Yang et al., "Wind power forecasting using deep learn-ing and hybrid features," Appl. Energy, vol. 210, pp. 53–65, 2018, doi: 10.1016/j.apenergy.2017.09.091.
[16] A. Elattar et al., "Weekly load forecasting using deep learn-ing," Electr. Power Syst. Res., vol. 186, 106409, 2020, doi: 10.1016/j.epsr.2020.106409.
[17] J. Schmidhuber, "Deep learning in neural networks: An overview," Neural Netw., vol. 61, pp. 85–117, 2015, doi: 10.1016/j.neunet.2014.09.003.
[18] M. T. Hagan, H. B. Demuth, y M. Beale, Neural Network Design, 2nd ed. PWS Publishing, 2014.
[19] S. Hochreiter y J. Schmidhuber, "Long short-term memory," Neural Comput., vol. 9, no. 8, pp. 1735–1780, 1997, doi: 10.1162/neco.1997.9.8.1735.
Licencia
Derechos de autor 2025 Matias Ezequiel Tielli, Franco Vega, Vanesa Hetze, Mario Orlando Blume
Esta obra está bajo una licencia internacional Creative Commons Atribución 4.0.
