A Temporal-Consistency-Enhanced BiLSTM Method for Photovoltaic Power Forecasting in Distributed PV Stations

Abstract

Distributed photovoltaic stations are increasingly operated as dense, weather-sensitive generation clusters, yet their short-term power forecasts often contain discontinuous ramps that are inconsistent with plant physics and recent temporal evolution. This paper proposes a temporal-consistency-enhanced bidirectional long short-term memory method, termed TC-BiLSTM, for 15-min-ahead to 4-h-ahead photovoltaic power forecasting in distributed PV stations. The method integrates normalized station measurements, meteorological variables, solar-position descriptors, and historical ramp features into a two-stream BiLSTM encoder. A compact consistency module then constrains adjacent forecast horizons by combining ramp-aware smoothness, irradiance-response monotonicity, and capacity-normalized error terms. Experiments are organized on a multi-station dataset containing 38 distributed PV stations, 18 months of operation, 96 daily sampling points, and four weather categories. Compared with persistence, SVR, CNN-LSTM, vanilla LSTM, and standard BiLSTM baselines, TC-BiLSTM reduces RMSE from 8.74% to 6.58% of installed capacity, decreases MAE from 5.21% to 3.91%, and improves ramp-event F1-score from 0.704 to 0.823. The gain is most visible on partly cloudy days, where the average ramp error falls by 22.6%. The results indicate that temporal consistency is a practical regularization principle for distributed PV forecasting because it improves point accuracy while producing smoother and more dispatchable forecast trajectories.