The volatility and uncertainty of wind power often affect the quality of electric energy, the security of the power grid, the stability of the power system and the fluctuation of the power market. In this case, the research on wind power forecasting is of great significance for ensuring the better development of wind power grids and the higher quality of electric energy. Therefore, a lot of new forecasting methods have been put forward. In this paper, a new forecasting model based on convolution neural network and LightGBM is constructed. The procedure is shown as follows, firstly, we construct new feature sets by analyzing the characteristics of the raw data on the time series from the wind field and adjacent wind field. Secondly, the convolutional neural network (CNN) is proposed to extract information from input data, and the network parameters are adjusted by comparing the actual results. After that, in consideration of the limitations of the single-convolution model in predicting wind power, we innovatively integrated LightGBM classification algorithm at the model to improve the forecasting accuracy and robustness. Finally, compared with the existing support vector machines, LightGBM and CNN, the fusion model has better performance in accuracy and efficiency.

Due to the volatility and randomness of the photovoltaic power generation, it is difficult for traditional models to predict it accurately. To solve the problem, we established a model based on the self-attention mechanism and multi-task learning to predict the ultra-short-term photovoltaic power generation. First, we selected the data with the optimal timing length and input the data into the Encoder-Decoder network based on the self-attention. The validity of features extracted by the encoder was checked by the decoder. Then, we added a restriction to the middle layer of the Encoder-Decoder network to prevent the autoencoder from copying the input to the output mechanically. This condition is used to predict the photovoltaic power generation, so a multi-task learning model was established. Finally, to take full advantage of the features that are efficiently expressed and allow our main task, the prediction task, to learn some unique features autonomously, we proposed a step-by-step training method and have validated the effectiveness of this view in experiments. Through experimental contrast, it is found that compared with the Encoder-Decoder network based on CNN and LSTM, the performance of the proposed method has been increased by 14.82% and 8.09% respectively. The RMSE and MAE of the Encoder-Decoder model based on the self-attention mechanism using step-by-step training are 0.071 and 0.040 respectively.

Generative large language models' (LLMs) inference suffers from inefficiency because of the token dependency brought by autoregressive decoding. Recently, speculative inference has been proposed to alleviate this problem, which introduces small language models to generate draft tokens and adopts the original large language model to conduct verification. Although speculative inference can enhance the efficiency of the decoding procedure, we find that it presents variable resource demands due to the distinct computation patterns of the models used in speculative inference. This variability impedes the full realization of speculative inference's acceleration potential in current systems. To tackle this problem, we propose SpecPIM to accelerate speculative inference on the PIM-enabled system. SpecPIM aims to boost the performance of speculative inference by extensively exploring the heterogeneity brought by both the algorithm and the architecture. To this end, we construct the architecture design space to satisfy each model's disparate resource demands and dedicate the dataflow design space to fully utilize the system's hardware resources. Based on the co-design space, we propose a design space exploration (DSE) framework to provide the optimal design under different target scenarios. Compared with speculative inference on GPUs and existing PIM-based LLM accelerators, SpecPIM achieves 1.52×/2.02× geomean speedup and 6.67×/2.68× geomean higher energy efficiency.