<p>Supporting Information.</p>

The natural gas supply crisis triggered by the Russia–Ukraine conflict has laid bare the energy market’s extreme vulnerability in the face of geopolitical risk, highlighting the need for accurate multi-step gas price forecasting. However, most AI-based energy price studies have a gap: they focus on single-step prediction or homogeneous model comparisons, lacking analysis of performance degradation in multi-step dynamic frameworks. This study takes daily natural gas price data from the Henry Hub in the United States from 1997 to 2024 as the research object, constructs a multi-step prediction framework with a step size ranging from 1 to 4 days, and systematically compares the prediction performances of four artificial intelligence models: feedforward neural network, support vector machine, random forest, and long short-term memory network. The quantitative results show that, across all prediction cycles, the long short-term memory model has the lowest error rate. For example, in one-step forecasting, its Mean Absolute Percentage Error is 8.53%. Practically, the findings matter. Taking European governments facing natural gas shortages in the Russia-Ukraine conflict as an example, LSTM models can be used for multi-step prediction to forecast price fluctuations 2–4 days in advance, optimizing import reserve strategies to avoid supply disruptions; energy traders can use this to design robust futures arbitrage portfolios. In summary, the research provides a scientific basis and reference for government energy security policy-making and institutional investor trading.