This paper investigates the effectiveness of Neural Circuit Policies (NCPs) compared to Long Short-Term Memory (LSTM) networks in forecasting time series data for energy production and consumption in the context of predictive maintenance. Utilizing a dataset generated from the energy production and consumption data of a Tuscan company specialized in food refrigeration, we simulate a scenario where the company employs a 60 kWh storage system and calculate the battery charge and discharge policies to assess potential cost reductions and increased self-consumption of produced energy. Our findings demonstrate that NCPs outperform LSTM networks by leveraging underlying physical models, offering superior predictive maintenance solutions for energy consumption and production.