<![CDATA[This paper systematically evaluates the performance of an LSTM baseline model, along with four smoothing augmentation methods (Kalman, Laplace, Moving Average, Savitzky-Golay), under different normalization strategies (Min-Max and Z-Score) for multivariate time-series forecasting. Experiments were conducted on six publicly available datasets (electricity consumption, energy consumption, sensor data, household energy, Indian electricity, and Brazilian temperature), and model performance was comprehensively compared using three metrics: MAPE, RMSE, and R². Results indicate that Laplace smoothing achieved the best performance across five datasets, effectively reducing errors while maintaining high fit quality, demonstrating its advantage in handling highly volatile and noisy time-series data. However, in some instances, Laplace smoothing, along with MA and SG methods, may produce an “over-smoothing” effect, causing forecasts to lose sensitivity to spike fluctuations. The choice of normalization strategy is equally critical: Min-Max is more suitable for data with stable distributions, while Z-Score demonstrates greater advantages for data with large numerical ranges and significant volatility. Notably, in temperature datasets with small sample sizes and high volatility, complex smoothing methods actually degraded performance, making the baseline LSTM + Z-Score the optimal choice. However, the LSTM-Laplace model with Min-Max normalization achieves the best performance among the models. Overall, the study concludes that improving prediction performance relies not only on model architecture but also on optimizing data scale, distribution characteristics, and preprocessing strategies.]]>
Copyrights © 2026
