Scholarlink

Uranium is critical in the production of nuclear power and formulation of a clean energy policy. Our research objective is to apply deep learning models to predict uranium prices with higher precision. We benchmark a range of Long Short-Term Memory (LSTM) models to forecast uranium price returns over a 360-month period. Applying LSTM models, we consider 3 types of optimisers, 8 optimiser layer configurations, up to 3 steps ahead training and benchmark against 6 types of ARIMA models and 6 exponential smoothing models. Compared to price data, returns are noisy which accounts for the superior performance of LSTM models. ARIMA models perform well, but LSTM tops the forecast benchmarks on all 3 optimisers achieving on average RMSE of 0.07. To outperform the best ARIMA and LSTM forecast models, we introduce a novel ensemble Variational Mode Decomposition technique. VMD-LSTM model leads to an average RMSE of 0.034404, MAE of 0.027578 and MAPE of 1.1698. This is a significant forecast improvement of 54.39%, 53.35% and 53.47% over the leading ADAM, SGDM and RMSPROP optimiser-based LSTM models respectively.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Forecasting Uranium Prices Using LSTM and VMD-Based Ensemble Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Sasheendran Gopalakrishnakone

摘要

Uranium is critical in the production of nuclear power and formulation of a clean energy policy. Our research objective is to apply deep learning models to predict uranium prices with higher precision. We benchmark a range of Long Short-Term Memory (LSTM) models to forecast uranium price returns over a 360-month period. Applying LSTM models, we consider 3 types of optimisers, 8 optimiser layer configurations, up to 3 steps ahead training and benchmark against 6 types of ARIMA models and 6 exponential smoothing models. Compared to price data, returns are noisy which accounts for the superior performance of LSTM models. ARIMA models perform well, but LSTM tops the forecast benchmarks on all 3 optimisers achieving on average RMSE of 0.07. To outperform the best ARIMA and LSTM forecast models, we introduce a novel ensemble Variational Mode Decomposition technique. VMD-LSTM model leads to an average RMSE of 0.034404, MAE of 0.027578 and MAPE of 1.1698. This is a significant forecast improvement of 54.39%, 53.35% and 53.47% over the leading ADAM, SGDM and RMSPROP optimiser-based LSTM models respectively.