Concept drift refers to unpredictable changes in the statistical properties of target variables over time, which degrades model performance in time series forecasting. Therefore, it is crucial to detect concept drift accurately. Existing concept drift detection methods based on data distribution focus on changes in the overall data distribution of all variables within specific windows, which are inadequate for tackling local temporal and joint distribution drift of smaller time frames and subsets of variables. Therefore, a multi-scale concept drift detection method called TD-IVDM (Time Dependency-Inter Variable Dependency Method) is proposed, which captures both temporal and inter-variable dependencies. This method consists of a data packaging module, a feature extraction module and a threshold comparison module. In the feature extraction module, the representation learning network TS2Vec is improved to extract time dependencies and a multi-dimensional Kernel Density Estimation (KDE) method is used to capture inter-variable dependencies. The changes in the distribution are qualified by calculating the difference in weighted features. In the threshold comparison module, a task-adaptive triple-threshold comparison module is proposed to compare the calculated differential features with the thresholds and output the severity level of concept drift. The results of experiments on artificial and real-world datasets validate the effectiveness of our method in detecting concept drift in time series forecasting tasks.