Big Data Time Series Forecasting Using Pattern Sequencing Similarity

Time series forecasting in big data analytics is crucial for making decisions in a variety of fields. but faces challenges due to high dimensionality, non-stationarity, and dynamic patterns. Conventional approaches frequently produce inaccurate results because they are unable to capture sudden variations and intricate temporal connections. This study proposes a Multi-scale Dynamic Time Warping-based Hierarchical Clustering (MDTWbH) approach to improve forecasting accuracy and scalability. Multi-scale Dynamic Time Warping (MDTW) transforms time series data into multi-scale representations, preserving local and global patterns, while Hierarchical Clustering groups similar sequences for enhanced predictive performance. The proposed framework integrates data preprocessing, outlier detection, and missing value interpolation to refine input data. It employs Apache Hadoop and Spark for efficient big data processing. Long Short Term Memory (LSTM) is applied within each cluster for accurate forecasting, and accuracy, precision, recall, F1-score, MAE, and RMSE are used to assess the performance of the model. Experimental results on electricity demand, wind speed, and taxi demand datasets demonstrate superior performance compared to existing techniques. MDTWbH provides a scalable and interpretable solution for large-scale time series forecasting by efficiently capturing evolving temporal patterns.