Contents I. Introduction
Currently, the demand for energy is increasing worldwide, and electricity, as an important component of energy consumption [1], presents challenges such as difficult large-scale storage and significant demand fluctuations. Therefore, in the power system, it is essential to maintain dynamic balance in generation, distribution, transmission, and consumption of electricity. Load forecasting starts from the known characteristics of the power system itself, considering external factors like weather and economy, analyzes historical data to study potential patterns and variations among various factors, and predicts future load data in advance. Load forecasting [2] can be categorized into ultra-short-term, short-term, and medium-to-long-term forecasting based on time scales, with short-term load forecasting focusing on predicting load demands for the next few hours to days. Traditional forecasting methods [3] include regression analysis, grey model method, auto regressive integrated moving average model, while machine learning methods [4] encompass support vector regression, ridge regression, random forest, neural networks, among which LSTM [5] stands out as a commonly used method for time series data prediction. LSTM consists of forget, input, and output gates and has several variants, with GRU [6] being a simplified version. Studies at [7] have shown that coupling the input gate and forget gate or removing the peephole connections simplifies LSTM without significantly reducing performance. However, existing time series prediction methods mainly focus on point predictions, lacking the capability to quantify prediction uncertainties. In [8], a QRNN monotonic model was used to predict the range of photovoltaic output. The results indicated that quantile regression does not require specific assumptions about the data distribution and can provide comprehensive information. However, a simple fully connected layer has poor ability to capture the time series. Therefore, in [9], a QRLSTM method was used to analyze the uncertainty of load. In addition, mainstream forecasting methods often involve various degrees of model improvements but fail to consider different load trends at different times. [10] conducted clustering and identification of massive load data, establishing identification models for different user groups. In [11], the authors optimized neural networks using a clustering algorithm. Both articles categorized large amounts of clustered load data and separately trained different models. However, they did not address the issue of data imbalance within historical data for the same target. During model training, the importance of different samples varies, leading to lower accuracy metrics for extreme load data and an inability to evenly learn all features of historical data. To address these challenges, this paper proposes an FCM-QRMGM-MDWS (Fuzzy Clustering–Membership Degree Weighted Sum) model. It first conducts similar day analysis and sample weighting based on fuzzy clustering, then trains load prediction models for different similar days using QRMGM prediction methods, and finally calculates quantile regression results by weighting different models of similar days using membership degree weights. This model considers the varying importance of different samples during model training, accounts for load uncertainties, and utilizes minimal gate control units to enhance training speed of load prediction models for short-term forecasting without compromising accuracy.
