Time series data is widely used for decision-making and advanced analytics such as forecasting. However, the vast data volumes make storage challenging. Using lossy compression can save more space compared to lossless methods, but it can affect the forecasting accuracy. Understanding the impact of lossy compression on forecasting accuracy is a multifaceted challenge, necessitating experimental evaluation across various forecasting models, compression methods, and time series. This paper conducts such experimental evaluation by combining seven forecasting models, three lossy compression algorithms, and six datasets. By simulating a real-life scenario where forecasting models use lossy compressed data for prediction, we address three main research questions related to compression error and its effects on the time series characteristics and the forecasting models. The results show that the Poor Man’s Compression and Swing Filter lossy compression algorithms add less error than the Squeeze method as the error bound increases. Poor Man’s Compression provides the best balance between compression ratio and forecasting accuracy. Specifically, we obtained an average compression ratio of 13.65, 5.56, and 14.97 for PMC, SWING, and SZ with an average impact on forecasting accuracy of 5.56%, 3.3%, and 8.5%, respectively. An analysis of several time series characteristics shows that the maximum Kullback-Leibler divergence between consecutive windows in the time series is the best indicator of the impact of lossy compression on forecasting accuracy. Finally, our results indicate that simple models like Arima, are more resilient to lossy compression than complex deep learning models. The source code and data are available at https://github.com/cmcuza/EvalImpLSTS.

Modern wind turbines are equipped with wired high-quality sensors that produce high-frequency sensor data in the form of time series as shown in Figure 1 a. From working with multiple different practitioners, we have learned that relatively few but very long high-quality time series are produced. The time series are either univariate, i.e., have one value per timestamp, or multivariate, i.e., have multiple values per timestamp. Further, they are either regular, i.e., have a fixed time interval between consecutive data points, or irregular. Despite these differences, the volume and velocity of the time series that are being produced are generally major challenges. For example, if the sensors are sampled at 100Hz, a single park of 100 wind turbines generates more than 11 PiB of data each year [1]. The sensor data is collected by weak edge devices and then transferred to powerful cloud servers over a relatively slow connection as shown in Figure 2. However, it is infeasible to transfer and store the raw time series due to their volume and velocity. Renewable energy system installations use low-end commodity PCs on the edge, e.g., 4 CPU cores, 4 GiB RAM, and an HDD [1]. In addition, the bandwidth between the edge and the cloud can be as low as 0.5-5 Mbit/s [1]. Thus, practitioners use simple aggregates, e.g., 10-minute averages, which remove valuable outliers and fluctuations as shown in Figure 1b. To remedy this, practitioners want to use lossy compression with a per-value error bound (E) to collect more high-frequency time series and thus improve their analytics.

With the lowering costs of sensors, high-volume and high-velocity data are increasingly being generated and analyzed, especially in IoT domains like energy and smart homes. Consequently, applications that require accurate short-term forecasts and predictions are also steadily increasing. In this paper, we provide an overview of a novel end-to-end platform that provides efficient ingestion, compression, transfer, query processing, and machine learning-based analytics for high-frequency and high-volume time series from IoT. The performance of the platform is evaluated using real-world dataset from RES installations. The results show the importance of high-frequency analytics and the surprisingly positive impact of error bounded lossy compression on machine learning in the form of AutoML. For example, when detecting yaw misalignments in wind turbines, an improvement of 9% in accuracy was observed for AutoML models on lossy compressed data compared to the current industry standard of 10-minute aggregated data. Thus, these small-scale experiments show the potential of the platform, and larger pilots are planned.

The rapid increase of traffic data generated by different sensing systems opens many opportunities to improve transportation services. An important opportunity is to enable stochastic routing that computes the arrival time probabilities for each suggested route instead of only the expected travel time. However, traffic datasets typically have many missing values, which prevents the construction of stochastic speeds. To address this limitation, we propose the Stochastic Spatio-Temporal Graph Convolutional Network (SST-GCN) architecture that accurately imputes missing speed distributions in a road network. SST-GCN combines Temporal Convolutional Networks and Graph Convolutional Networks into a single framework to capture both spatial and temporal correlations between road segments and time intervals. Moreover, to cope with datasets with many missing values, we propose a novel self-adaptive context-aware diffusion process that regulates the propagated information around the network, avoiding the spread of false information. We extensively evaluate the effectiveness of SST-GCN on real-world datasets, showing that it achieves from 4.6% to 50% higher accuracy than state-of-the-art baselines using three different evaluation metrics. Furthermore, multiple ablation studies confirm our design choices and scalability to large road networks.