Levenberg-Marquardt-Optimized Feed-Forward Neural Network for High-Accuracy Day-Ahead Joint Forecasting of Electricity Load and Market Prices in Smart Grid Environments

The modern electrical grid is no longer just a one-way street where power flows from a giant plant to a passive home. It has become a dynamic, two-way system where solar panels on rooftops, electric vehicles, and smart appliances constantly interact with the market. In this complex environment, two numbers matter more than almost any others: how much electricity people will need tomorrow, and what that electricity will cost. These two figures are deeply linked; when demand spikes or the wind stops blowing, prices can shoot up, and when prices rise, people often change how they use power. Getting these predictions right is vital. If a utility company guesses wrong, it might waste money buying expensive power or leave the lights flickering. If a power plant guesses wrong, it might miss out on selling energy or get stuck with a bill it cannot pay. For decades, experts have tried to forecast these numbers using statistical tools, but the sheer unpredictability of weather and human behavior has made the task incredibly difficult.

A team of researchers has now proposed a new way to tackle this problem, focusing on a specific type of computer program designed to learn from past patterns. Instead of trying to predict the load and the price separately, they built a single system that guesses both at the same time. Think of this system like a student who studies a history book to understand how two related events, like the weather and the crowd size at a park, influence each other, rather than trying to learn about the weather and the crowd in two different classes. The researchers used a method called the Levenberg-Marquardt algorithm to train this system. In simple terms, this is a sophisticated mathematical technique that helps the computer learn much faster and more reliably than older methods, adjusting its internal settings to find the best possible fit for the data without getting stuck in dead ends.

The team tested their approach using real-world data from the PJM Interconnection, a massive electricity market that covers parts of the eastern United States. They fed the computer hours of historical information, including past electricity usage, past prices, temperature forecasts, and even predictions of how much renewable energy would be available. The computer was asked to look at all this information and predict the next twenty-four hours of both load and price simultaneously. The results were impressive. When compared to other popular forecasting tools, including standard neural networks and traditional statistical models, this new system consistently made fewer mistakes for both the electricity load and the market prices. It achieved the lowest error metrics across the board, outperforming baseline methods like persistence and ARIMA, and demonstrating that the joint modeling approach successfully captured the complex dynamics of the market.

What makes this finding particularly useful is not just the accuracy, but the speed and simplicity of the solution for load forecasting. Many modern artificial intelligence models require massive amounts of computing power and take a long time to train, making them hard to use in real-time situations where decisions must be made quickly. This new model, however, is relatively compact and trains very quickly. It manages to capture the complex, hidden relationships between how much power people use and what that power costs, without needing the heavy machinery of more complicated deep-learning systems. The researchers found that by sharing the "thinking" part of the network between the two predictions, the system learned that high demand often leads to high prices, and that these patterns repeat in predictable ways based on the time of day and the day of the week.

The study also looked closely at how the model performed over time, checking if it could handle different seasons and unusual market conditions. The analysis showed that the errors for load were spread out evenly, meaning the system did not consistently overestimate or underestimate the numbers. It handled the daily cycles well, predicting the morning and evening peaks in usage with high fidelity. While the researchers noted that the accuracy of the price prediction still depends heavily on the quality of the weather and renewable energy forecasts provided to the system, the core method proved robust for both variables. They also pointed out that this approach focuses on giving a single best guess for the future, rather than a range of possibilities, which is a limitation for some risk-management tasks but ideal for immediate operational planning.

Ultimately, this work demonstrates that sometimes the most effective solution is not the most complex one. By refining an older, well-understood type of computer program with a powerful training method, the team created a tool that is both fast and accurate enough to help manage the delicate balance of a modern smart grid regarding electricity demand and market prices. The ability to predict both the demand for electricity and its cost in a single, quick calculation offers a practical advantage for the people who keep the lights on and the markets running. As the grid continues to evolve with more renewable energy and electric vehicles, having a reliable, efficient way to anticipate these twin challenges becomes increasingly essential. This study suggests that with the right mathematical tools, we can navigate the uncertainty of the future grid with a clearer view than ever before.