Time series prediction has been realized in numerous real-world applications. Many existing studies mainly focus on general trend prediction by minimizing the prediction error on the whole time series. However, predicting peak values (i.e., the values are much higher than average of the time series) such as peak traffic volumes and network traffic, is more important for effective congestion control. The prediction of peak values is challenging due to the sparsity of the peak values. To tackle the data imbalance issue, we propose a novel model which utilizes Multi-Task Learning (MTL) for univariate and short-term peak value prediction. Specifically, we decompose the task of peak value prediction into two sub-tasks: (1) the regression task which aims at predicting the value of the targeted time step, and (2) the classification task which aims at classifying whether the targeted time step is a peak. A novel loss function, namely Peak Weighted Quantile Loss (PWQL), is proposed to improvement the performance of peak value prediction by giving more penalties on underestimating peak values. The experimental results on real-world traffic data show that our model can achieve prediction performance close to the state-of-the-art peak value prediction model using only around 40% parameters, showing greater potential for deployment in resource-constrained environments.