The default Horizontal Pod Autoscaler strategy in Kubernetes has limitations due to its inherent response mechanism, leading to scaling delays. To improve resource response performance and resource utilization, this paper introduces an elastic scaling strategy based on time-series resource load prediction. The proposed prediction model, WTT-iTransformer, is specifically designed to forecast cluster resources. It is known that iTransformer excels in long-term sequence prediction and can capture the correlations between multiple variables by embedding variable sequences as tokens. By adding a Wavelet Transform Convolutional layer and integrating a Multi-Scale Temporal Convolutional Network, the WTT-iTransformer model is constructed, enabling more precise extraction of long-term features and dependencies in resource time-series from both the time and frequency domains, which aligns better with the prediction of container usage characteristics. Based on the load variation prediction of this model, rapid scaling can be implemented at the early stages of high and low traffic occurrences, addressing the issues of delayed responses and low resource utilization. Experimental results show that the WTT-iTransformer demonstrates better stability and lower training error during training, accurately predicting cluster load trends. The improved elastic scaling strategy, compared to the traditional HPA in Kubernetes, is more intelligent and stable, showing significant improvements in scenarios with notable load characteristics and frequent burst traffic, thus holding broad application potential.