Long-sequence multivariate time series forecasting is widely applicable across various real-world scenarios. The Vanilla transformer has recently been used to predict the long-sequence time series. Transformers are initially designed for tasks like machine translation, where the order of the input sequence is crucial, but the specific temporal dependencies of the time series may not be fully captured. Moreover, its autoregressive nature poses limitations. This paper proposes a novel approach named FANT (frequency-attention-based non-autoregressive transformer) tailored for accurately forecasting long-sequence multivariate time series. The FANT method decomposes the time series into trend and seasonality components and enriches the input data with a calculated time vector using TimetoVec. This augmentation aids in capturing both short-term and long-term dependencies within the time series, facilitating the model’s comprehension of intricate temporal relationships. The approach employs the Fourier transform to shift the data into the frequency domain, where self-attention is subsequently calculated to discern patterns and periodicity within the data. Importantly, the FANT model is non-autoregressive, enabling simultaneous generation of all sequence elements. This characteristic minimises error propagation and reduces computational time, enhancing overall efficiency. The empirical evaluation of the FANT model encompasses seven datasets from financial and energy domains. Results indicate its superiority over the vanilla Transformer, showcasing a minimum relative improvement of 18.74% across all datasets. Additionally, an ablation study underscores the effectiveness of the frequency-attention model compared to self-attention.