Abstract: Abstract: Short-term load forecasting plays a central role in the daily operation and dispatching of power systems. Greater flexibility and uncertainty in the operation of power systems have brought harsh requirements on the accurate forecasting of short-term load, particularly with the in-depth advancement of power market reform. A high-precision prediction model is also highly demanding to effectively coordinate the relationship between power generation, transmission, distribution, and consumption. In this study, short-term load forecasting was therefore proposed using variational mode decomposition (VMD) and particle swarm optimization (PSO) modal combination. The VMD was adopted for the adaptive signal decomposition of load sequence, considering the time and frequency domain in the signal decomposition evaluation. The impact of two dimensions on authenticity, independence, and performance was quantitatively clarified, further to determine the evaluation indicators of signal decomposition. An authenticity test of signal decomposition included the redundant component and residual difference component. In redundant components, the Pearson correlation coefficient of the component and the original signal was compared with the authenticity threshold from the perspective of the time domain. In residual components, the ratio of frequency band with significant amplitude in the residual spectrum to the original signal frequency band was used to measure the spectral characteristics of residual components from the perspective of the frequency domain. In the independence index test, two indicators were used to evaluate in the two dimensions of time and frequency domain, including the average Pearson correlation coefficient between the signal components, and the average overlap of significant frequency bands. Time domain was measured using the average Pearson correlation coefficient of each component and the original signal. The average concentration index of the significant frequency band was also selected to verify the signal decomposition. Three indicators were integrated to establish a comprehensive evaluation indicator for signal decomposition. PSO was then used to optimize the parameters of the VMD model, where the comprehensive evaluation index of signal decomposition was taken as the objective function. The influencing factors of daily load included temperature, humidity, historical load, and day-ahead electricity price. The modal combination was used to integrate the modal sequence of influencing factors with similar periodicity in the frequency domain of load. The number of influencing factors after combination was expanded the same as the significant frequency bands of the load frequency domain, where the combined influencing factor modal presented a strong correlation with the load sequence. The expanded variables of influencing factors were input into the PSO-least squares support vector machine (LSSVM) model for the load forecast of the day. The simulation results show that the VMD decomposition using the evaluation index optimization was better than the wavelet analysis, optimized ensemble empirical mode decomposition (EEMD), and non-optimized VMD, indicating the decomposed modal sequence behaved higher quality. Specifically, The maximum relative error increased by 3.36 percentage point, the average relative error increased by 1.71 percentage point, the maximum absolute error increased by 95MW, and the average absolute error increased by 55.72 MW, compared with the PSO-LSSVM using a modal combination. This finding can provide sound support to the construction of a power market with flexible power regulation under the penetration of a high proportion of renewable energy, efficient grid interconnection, and extensive user response.