The efficient management of electrical energy consumption represents a critical challenge in developing regions, particularly in Iraq where infrastructure limitations and fluctuating supply necessitate innovative monitoring solutions. This paper presents a comprehensive study on the design, implementation, and evaluation of a cost-effective Internet of Things (IoT)-based smart metering system deployed across five residential sites in Musayyib District, Babel Governorate, Iraq. The system integrates PZEM-004T energy sensors with NodeMCU ESP8266 development boards to enable real-time collection of electrical parameters including voltage, current, active power, energy consumption, frequency, and power factor. The collected data serves as the foundation for developing and comparing six distinct machine learning algorithms for energy consumption forecasting: Linear Regression, Autoregressive models, Artificial Neural Networks, Convolutional Neural Networks, XGBoost, and Random Forest. The experimental results reveal significant challenges in achieving high forecasting accuracy within this specific operational context, with most models exhibiting low or negative R2 values and elevated Mean Absolute Percentage Errors. However, tree-based ensemble methods, particularly XGBoost and Random Forest, demonstrated relatively superior performance compared to linear and basic neural network approaches. This research contributes to the understanding of practical IoT deployment challenges in resource-constrained environments and emphasizes the critical importance of comprehensive feature engineering, data quality assurance, and appropriate model selection for developing reliable energy forecasting systems in similar contexts.