Travel time duration and traffic forecasting in urban environments is of critical importance for efficient supply chain logistics and accurate navigation services. The paper compares data-driven ...statistical and machine learning approaches for urban traffic travel times forecasting on a 15-minute resolution for the next 24 hours. A detailed analysis of the available historical data is performed, including historical floating car data for a one-year period. Data analysis is followed by development and evaluation of baseline, ARIMAX, neural network and gradient boosting models. The predictions are generated on a day-ahead forecasting horizon where the machine learning models are propagated in a recursive strategy. Recursive approach implies more accurate short-term forecasts that are crucial in the context of nowcasting, while maintaining good accuracy on long-term horizons necessary for daily operation planning. The evaluation of the models accuracy is conducted on four main arterial urban routes in the city of Zagreb, Croatia, with travel times data obtained from a commercial navigation service. A detailed analysis of the results is performed, with a special focus put on the accuracy of near future predictions, as well as different models and model structures.
Enhanced accuracy and long-term predictions of ship motion during sea operations can effectively mitigate safety risks associated with aircraft takeoff and landing on board.This paper proposes a ...transformer-based ship motion attitude prediction model.Our work leverages a novel self-attention mechanism with adaptive position encoding and learnable attention weights to improve long-term prediction accuracy. Furthermore, we also incorporate a pretraining phase using a random masking strategy to enhancing the model's training capability and reducing prediction phase duration.The proposed model is evaluated using data from a ship undergoing constant speed and Z-word motion to predict the roll and pitch angles of the ship. The model is compared with ARMA, EMD-ARMA, LSTM, Bi-LSTM and traditional Transformer models. The experimental results demonstrate that the proposed method outperforms these models in multi-step prediction scenarios.
Lithium-ion battery temperature monitoring contributes to the higher performance of lithium batteries and reduces the risk of thermal runaway. Since the battery temperature can be approximated as a ...time series, this work reports a new model named convolutional transformer (Convtrans) for multi-step time series forecasting, which obtains pleasing results. To evaluate our model, we present a cross-sectional comparison of the model with the other three mainstream algorithms in multi-step time series forecasting and a vertical comparison with single-step time forecasting. On one hand, Convtrans has the minimum root mean square error with the highest prediction accuracy and can also predict the trend and shape of temperature curves compared with the other three mainstream algorithms, which means that it has the best results in multi-step time series forecasting. On the other hand, compared with single-step time series forecasting, Convtrans predicts 24 times temperature data while doesn't sacrifice much accuracy even though it costs 6 times running time. Furthermore, in the case of predicting more points, it also maintains good accuracy and can perfectly predict trends of the temperature. In all, we prove the superiority of multi-step time series forecasting over a long period. Therefore, multi-step time series forecasting based on Convtrans can sever as the battery temperature prognostic technology providing timely warnings to assist battery thermal management.
Display omitted
•Convtrans algorithm is developed for multi-step time series forecasting of battery temperatures.•Compared with other multi-step prediction algorithms, Convtrans has smaller prediction error.•Compared with single-step time series forecasting, Convtrans can predict more temperature data without losing accuracy.•Convtrans can provide timely warnings to assist battery thermal management.
Time series forecasting is ubiquitous in various scientific and industrial domains. Powered by recurrent and convolutional and self-attention mechanism, deep learning exhibits high efficacy in time ...series forecasting. However, the existing forecasting methods are suffering some limitations. For example, recurrent neural networks are limited by the gradient vanishing problem, convolutional neural networks cost more parameters, and self-attention has a defect in capturing local dependencies. What’s more, they all rely on time invariant or stationary since they leverage parameter sharing by repeating a set of fixed architectures with fixed parameters over time or space. To address the above issues, in this paper we propose a novel time-variant framework named Self-Attention-based Time-Variant Neural Networks (SATVNN), generally capable of capturing dynamic changes of time series on different scales more accurately with its time-variant structure and consisting of self-attention blocks that seek to better capture the dynamic changes of recent data, with the help of Gaussian distribution, Laplace distribution and a novel Cauchy distribution, respectively. SATVNN obviously outperforms the classical time series prediction methods and the state-of-the-art deep learning models on lots of widely used real-world datasets.
