Embedded automation and control systems increasingly depend on continuous monitoring of environmental variables, particularly temperature and relative humidity, under strict energy and computational constraints. Recent advances in Tiny Machine Learning (TinyML) enable predictive models to be executed directly on microcontrollers, requiring explicit trade-offs between predictive accuracy, memory footprint, execution latency, energy consumption, and operational robustness. This work presents a comparative evaluation of three lightweight neural network architectures—a multilayer perceptron (MLP), a one-dimensional convolutional neural network (Conv1D Tiny), and a long short-term memory network (LSTM)—implemented on an ESP32 microcontroller for temperature and humidity time-series modeling. Two execution scenarios are investigated, in which both replay and field modes employ the same on-device rolling window composed of 24 valid samples. In replay mode, deterministic input data are used as a deterministic test bench for controlled validation. In field mode, the rolling window advances as new sensor samples are acquired during real operation. Experiments were conducted using an offline evaluation workflow, referred to as LiteML-Edge, employed as an experimental tool for model training, testing, and consistency checks between offline evaluation and on-device execution. Model performance is assessed using energy-aware deployment-orientated criteria central to control systems, including inference latency, flash and RAM utilization, and energy-related measurements, together with standard regression metrics such as mean absolute error (MAE), root mean squared error (RMSE), and coefficient of determination (R²). Results indicate that the LSTM achieves higher predictive accuracy under controlled replay conditions, while the MLP demonstrates higher robustness and lower computational overhead during field operation. The Conv1D Tiny model exhibits intermediate behavior, balancing limited temporal modeling capability with moderate memory usage and energy efficiency. These results confirm that no single architecture is universally optimal and that model selection should be guided by execution context and control constraints.