Currently

PV power forecasting methods pursue the highest overall forecast accuracy as their optimized goal

thus the forecasting power may be higher than the actual power during some time intervals. The generation scheduling plan formulated on the basis of overestimated forecasting power may not meet the actual power demand

affecting the reliable power supply. To solve this problem

a forecasting method for provincial PV intra-day guaranteed power output with the goal of ensuring safety is proposed. Firstly

the spatial feature extraction abilities of convolutional neural networks (CNN) and the time-series processing abilities of bi-directional long short-term memory (BiLSTM) were utilized

and a provincial PV intra-day guaranteed power output forecasting model based on CNN-BiLSTM was constructed. A multi-objective loss function was introduced

which includes accuracy of prediction of the positive deviation in forecasting power

the degree of forecasting power correction

recognition of positive deviation

and model response speed

to identify and extract the temporal features of positive deviation of PV forecasting power. Two evaluation indexes

namely

the guaranteed rate and the guaranteed energy ratio

are established for PV intra-day guaranteed power output forecasting results. The effectiveness of the proposed method is verified based on the actual production and operation data of a province in northwest China.