<![CDATA[In the automotive industry, forecasting future demand is particularly crucial due to the complexity of production processes and supply chains. This article examines the comparative performance of a hybrid CNN-LSTM model for car sales forecasting, utilizing seven optimization algorithms: Adam, RMSprop, SGD, Adagrad, Adadelta, Adamax, and Nadam. Each optimization method has its own advantages. For instance, Adam offers fast convergence, while RMSprop is more effective in handling large gradient fluctuations. Adagrad is well-suited for managing gradient magnitude variations, whereas Adadelta addresses Adagrad’s limitations. Adamax is ideal for models with a broader parameter space, and Nadam combines Nesterov Accelerated Gradient and Adam, making it suitable for tasks requiring both momentum and adaptive learning. This study demonstrates that the CNN-LSTM model optimized with Nadam delivers the best performance, achieving a Mean Squared Error (MSE) of 35,383.14 and a Root Mean Squared Error (RMSE) of 188.10. In comparison, traditional methods such as ARIMA yield an MSE of 59,105.94 and an RMSE of 243.11. These findings indicate that the CNN-LSTM model optimized with Nadam outperforms conventional time series forecasting methods in predictive accuracy.]]>
Copyrights © 2025
